CROSS REFERENCE TO RELATED APPLICATIONS

This application is a continuation of U.S. Ser. No. 15/486,102 filed Apr. 12, 2017, the entire content of which is incorporated herein by reference. U.S. Ser. No. 15/486,102 is a continuation application of International Patent Application No. PCT/CN2016/077280, filed on Mar. 25, 2016, which claims priority to Chinese Patent Application No. 201510154027.0, filed on Apr. 2, 2015.

TECHNICAL FIELD

The present disclosure generally relates to the field of image recognition, and more particularly, to a method and device for training for a CNN model.

BACKGROUND

In the field of image recognition, a CNN model is usually adopted to determine a classification of an image to be recognized. Before the classification of the image to be recognized is recognized through the CNN model, the CNN model is required to be trained at first.

Training of a CNN model is usually implemented in a manner as follows. At first, model parameters of the CNN model to be trained are initialized, the model parameters including initial convolution kernels of the respective convolution layer, initial bias matrixes of the respective convolution layer, and an initial weight matrix and an initial bias vector of a fully connected layer. Then, an area to be processed with a fixed height and a fixed width is acquired from each of preselected training images, the fixed height and the fixed width being matched with a classification of an image to be recognized, which is preset, as an image that can be processed, by the CNN model to be trained. The area to be processed corresponding to each of the training images is input into the CNN model to be trained. Next, on each convolution layer, convolution operation and maximal pooling operation are performed on each area to be processed by using the initial convolution kernel and initial bias matrix of each convolution layer, to obtain a feature image of each area to be processed on each convolution layer. Then, each feature image is processed to obtain classification probability of each area to be processed by using the initial weight matrix and initial bias vector of the fully connected layer. Then, a classification error is calculated according to initial classification and the classification probability of each of the training images. A mean of the classification errors is calculated according to the classification errors of all the training images. Then, the model parameters of the CNN model to be trained are regulated by using the mean of the classification errors. Then, the abovementioned respective steps are iterated for a specified number of times by using the regulated model parameters and the respective training images. Finally, model parameters obtained when the number of iterations reaches the specified number of times are determined as the model parameters of the trained CNN model.

In a process of implementing the present disclosure, the inventor finds that a related technology at least has a problem as follows.

It is necessary to acquire the areas to be processed with the fixed height and the fixed width from the pre-selected training images in the process of training a CNN model, and accordingly the trained CNN model can only recognize an image with the fixed height and the fixed width, which may cause certain limits to image recognition of the trained CNN model and further cause a limited application range.

SUMMARY

In order to solve the problem of a conventional art, the embodiments of the present disclosure provide a method and device for training a CNN model. The technical solutions are implemented as follows.

In a first aspect, a method for training a CNN model is provided, which may include:

initial model parameters of a CNN model to be trained are acquired, the initial model parameters including initial convolution kernels and initial bias matrixes of convolution layers of respective levels, and an initial weight matrix and an initial bias vector of a fully connected layer;

a plurality of training images are acquired;

on the convolution layer of each level, convolution operation and maximal pooling operation are performed on each of the training images to obtain a first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix of the convolution layer of each level;

horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain a second feature image of each of the training images on the convolution layer of each level;

a feature vector of each of the training images are determined according to the second feature image of each of the training images on the convolution layer of each level;

each feature vector is processed to obtain a classification probability vector of each of the training images according to the initial weight matrixes and the initial bias vectors;

a classification error is calculated according to the classification probability vector and initial classification of each of the training images;

the model parameters of the CNN model to be trained are regulated on the basis of the classification errors;

on the basis of the regulated model parameters and the plurality of training images, the process of regulating the model parameters is continued until the number of iterations reaches a preset number; and model parameters obtained when the number of iterations reaches the preset number are determined as the model parameters of the trained CNN model.

In a second aspect, a device for training a CNN model is provided, which may include:

a first acquisition module configured to acquire initial model parameters of a CNN model to be trained, the initial model parameters including initial convolution kernels and initial bias matrixes of convolution layers of respective levels, and an initial weight matrix and an initial bias vector of a fully connected layer;

a second acquisition module configured to acquire a plurality of training images;

a feature extraction module configured to, on the convolution layer of each level, perform convolution operation and maximal pooling operation on each of the training images to obtain a first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix of the convolution layer of each level;

a horizontal pooling module configured to perform horizontal pooling operation on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain a second feature image of each of the training images on the convolution layer of each level;

a first determination module configured to determine a feature vector of each of the training images according to the second feature image of each of the training images on the convolution layer of each level;

a processing module configured to process each feature vector to obtain a classification probability vector of each of the training images according to the initial weight matrixes and the initial bias vectors;

a calculation module configured to calculate a classification error according to the classification probability vector and initial classification of each of the training images;

a regulation module configured to regulate the model parameters of the CNN model to be trained on the basis of the classification errors;

an iteration module configured to, on the basis of the regulated model parameters and the plurality of training images, continue the process of regulating the model parameters until the number of iterations reaches a preset number; and a second determination module configured to determine model parameters obtained when the number of iterations reaches the preset number as the model parameters of the trained CNN model.

In a third aspect, a server is provided, which may include:

one or more processors, and

a memory connected with the one or more processors, the memory being configured to store instructions executable for the one or more processors,

wherein the one or more processors may be configured to execute the instructions stored by the memory to perform the method for training the CNN model provided on the first aspect.

The technical solutions provided by the embodiments of the present disclosure can achieve the following beneficial effects.

After the convolution operation and the maximal cooling operation are performed on the training images on the convolution layer of each level, the horizontal pooling operation is further performed on the feature images obtained by the maximal cooling operation. During the horizontal cooling operation, the feature images which identify features of the feature images in a horizontal direction may further be extracted from the feature images to ensure that the trained CNN model may break limits of a size of an image to be recognized in the horizontal direction during image classification recognition and recognize an image to be recognized with any size, so that an image recognition application range of the CNN model trained in such a manner is relatively wider.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to describe the technical solutions in the embodiments of the present disclosure more clearly, the accompanying drawings which are required by description about the embodiments will be simply introduced below. Apparently, the accompanying drawings described below are only some embodiments of the present disclosure, and those skilled in the art may further obtain other accompanying drawings according to these accompanying drawings without creative work.

FIG. 1 is a flow chart showing a method for training a CNN model, according to an embodiment of the present disclosure.

FIG. 2 is a flow chart showing a method for training a CNN model, according to another embodiment of the present disclosure.

FIG. 3 is a schematic diagram illustrating a CNN model to be trained, according to another embodiment of the present disclosure.

FIG. 4 is a schematic diagram illustrating a process of a horizontal pooling operation, according to another embodiment of the present disclosure.

FIGS. 5A-5B are flow charts showing a method for training a CNN model, according to another embodiment of the present disclosure.

FIG. 6 is a schematic diagram of a structure of a device for training a CNN model, according to another embodiment of the present disclosure.

FIG. 7 is a schematic diagram of a structure of a terminal, according to another embodiment of the present disclosure.

FIG. 8 is a schematic diagram of a structure of a server, according to another embodiment of the present disclosure.

DETAILED DESCRIPTION

In order to make a purpose, technical solutions and advantages of the present disclosure clearer, implementations of the present disclosure will be further described in detail below with reference to the accompanying drawings.

FIG. 1 is a flow chart showing a method for training a CNN model, according to an embodiment of the present disclosure. As shown in FIG. 1, a flow of the method provided by the embodiments of the present disclosure includes the following steps.

In Step 101, initial model parameters of a CNN model to be trained are acquired, wherein the initial model parameters include initial convolution kernels and initial bias matrixes of convolution layers of respective levels, and an initial weight matrix and an initial bias vector of a fully connected layer.

In Step 102, a plurality of training images are acquired.

In another embodiment, the step that the plurality of training images are acquired includes:

a plurality of initial training images are acquired;

for each of the initial training images, a width-height ratio of the initial training image is kept, and the initial training image is processed to obtain a first image with a specified height; and

the first image is processed to obtain a second image with a specified width, and the image with the specified height and the specified width is determined as the training images corresponding to the initial training image.

In another embodiment, the step that the first image is processed to obtain the second image with the specified width includes:

when width of the first image is smaller than the specified width, left and right sides of the first image are uniformly filled with pixels having a specified gray-scale value, and the second image is obtained when the width of the first image reaches the specified width; and

when the width of the first image is larger than the specified width, pixels on the left and right sides of the first image are uniformly cropped, and the second image is obtained when the width of the first image reaches the specified width.

In another embodiment, the step that the plurality of training images are acquired includes:

a plurality of initial training images are acquired; and

for each of the initial training images, a width-height ratio of the initial training image is kept, the initial training image is processed to obtain an image with a specified height, and a width corresponding to the specified height is determined as width of the initial training image.

In another embodiment, the training images are images in a natural scene, the images in the natural scene include characters in different languages, and the CNN model to be trained is a language recognition classifier.

In Step 103, on the convolution layer of each level, convolution operation and maximal pooling operation are performed on each of the training images to obtain a first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix of the convolution layer of each level.

In another embodiment, the step that the convolution operation and the maximal pooling operation are performed on each of the training images to obtain the first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix on the convolution layer of each level includes:

for each of the training images, the first feature image on the convolution layer of a previous level is input to a current convolution layer, and the convolution operation is performed on the first feature image on the convolution layer of the previous level to obtain a convolutional image on the current convolution layer by using the initial convolution kernel and initial bias matrix of the current convolution layer, wherein the first feature image on the convolution layer of the previous level is the training image if the current convolution layer is the convolution layer of the first level; and

after the maximal pooling operation is performed on the convolutional image on the current convolution layer to obtain the first feature image of the training image on the current convolution layer, the first feature image on the current convolution layer is continuously transmitted to the convolution layer of a next level, and the convolution operation and the maximal pooling operation are performed on the convolution layer of the next level until the convolution operation and the maximal pooling operation are performed on the convolution layer of a last level to obtain the first feature image on the convolution layer of the last level.

In Step 104, horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain a second feature image of each of the training images on the convolution layer of each level.

In another embodiment, the step that the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain the second feature image of each of the training images on the convolution layer of each level includes:

for the first feature image of each training image on the convolution layer of each level, a maximum value of elements of each of rows of each of the images in the first feature image on the convolution layer is extracted, wherein the first feature image includes a preset number of images, and the preset number is the same as each of the numbers of the convolution kernels and bias matrixes of the convolution layer;

the maximum values extracted from all the rows of each image are arranged into a one-dimensional vector according to arrangement of pixels of each image; and

the one-dimensional vectors of all the images in the first feature image on the convolution layer are combined to obtain the second feature image on the convolution layer.

In Step 105, a feature vector of each of the training images is determined according to the second feature image of each of the training images on the convolution layer of each level.

In another embodiment, the step that the feature vector of each of the training images is determined according to the second feature image of each of the training images on the convolution layer of each level includes:

for each of the training images, elements of all rows of the second feature image of the training image on the convolution layer of each level are connected head to tail to obtain the feature vector of the training image.

In Step 106, each feature vector is processed to obtain a classification probability vector of each of the training images according to the initial weight matrixes and the initial bias vectors.

In Step 107, a classification error is calculated according to the classification probability vector and initial classification of each of the training images.

In another embodiment, the step that the classification error is calculated according to the classification probability vector and initial classification of each of the training images includes:

the initial classification of each of the training images are acquired;

the classification error of each of the training images is calculated according to the classification probability vector and initial classification of each of the training images by using the following formula:

Loss=−Lny_label,

where Loss represents the classification error of each of the training images, label represents the initial classification of each of the training images, y_irepresents a element of the classification probability vector of each of the training images, and y_labelrepresents a classification probability corresponding to the initial classification; and

the mean of the classification errors of all the training images is calculated, and the mean of the classification errors is determined as a classification error.

In Step 108, the model parameters of the CNN model to be trained are regulated on the basis of the classification errors.

In Step 109, on the basis of the regulated model parameters and the plurality of training images, the process of regulating the model parameters is continued until the number of iterations reaches a preset number.

In Step 110, model parameters obtained when the number of iterations reaches the preset number are determined as the model parameters of the trained CNN model.

According to the method provided by the embodiments of the present disclosure, after the convolution operation and the maximal pooling operation are performed on the training images on the convolution layer of each level, the horizontal pooling operation is further performed on the feature images obtained by the maximal cooling operation. During the horizontal cooling operation, the feature images which identify the features of the feature images in a horizontal direction may further be extracted from the feature images to ensure that the trained CNN model may break limits of a size of an image to be recognized in the horizontal direction during image classification identification and recognize an image to be recognized with any size, so that an image recognition application range of the CNN model trained in such a manner is relatively wider.

In another embodiment, the CNN model to be trained includes four levels of convolution layers and two fully connected layers, and the convolution layers of the respective levels include the same or different numbers of convolution kernels and bias matrixes.

The step that the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain the second feature image of each of the training images on the convolution layer of each level includes:

the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of a second level, the first feature image of each of the training images on the convolution layer of a third level and the first feature image of each of the training images on the convolution layer of a fourth level to obtain the second feature image of each of the training images on the convolution layer of the second level, the second feature image of each of the training images on the convolution layer of the third level and the second feature image of each of the training images on the convolution layer of the fourth level, respectively.

The step that the feature vector of each of the training images is determined according to the second feature image of each of the training images on the convolution layer of each level includes:

for each of the training images, the feature vector of the training image is determined according to the second feature image of the training image on the convolution layer of the second level, the second feature image of the training image on the convolution layer of the third level and the second feature image of the training image on the convolution layer of the fourth level.

All of the optional technical solutions may be freely combined into optional embodiments of the present disclosure, which will not be elaborated one by one.

With reference to a content of the embodiment corresponding to FIG. 1, FIG. 2 is a flow chart showing a method for training a CNN model, according to another embodiment of the present disclosure. As shown in FIG. 2, a flow of the method provided by the embodiment of the present disclosure includes the following steps.

In Step 201, initial model parameters of a CNN model to be trained are acquired, wherein the initial model parameters include initial convolution kernels and initial bias matrixes of convolution layers of respective levels, and an initial weight matrix and an initial bias vector of a fully connected layer.

Specifically, a process of training the CNN model is a process of determining model parameters of the CNN model to be trained. When the model parameters is determined, the initial model parameters may be initialized for the CNN model to be trained. The initial model parameters are continuously optimized in a subsequent training process. And an optimal model parameters obtained by optimization are determined as the model parameters of the trained CNN model. Therefore, when the CNN model to be trained is trained, it is necessary to acquire the initial model parameters of the CNN model to be trained at first.

In the case, the CNN model to be trained usually includes at least two levels of convolution layers and at least one level of fully connected layer, the convolution layer of each level includes a plurality of convolution kernels and a plurality of bias matrixes, and each level of fully connected layer includes a plurality of weight matrixes and a plurality of bias vectors. Therefore, the acquired model parameters include the initial convolution kernels of the convolution layer of each level, the initial bias matrixes of the convolution layer of each level, the initial weight matrix of the fully connected layer and the initial bias vector of the fully connected layer.

The number of the convolution layers and the number of the fully connected layers of the CNN model to be trained are not specifically limited in the embodiment of the present disclosure, and may be set according to a requirement during specific implementation. For example, FIG. 3 shows a schematic diagram illustrating a CNN model to be trained. The CNN model to be trained shown in FIG. 3 includes four levels of convolution layers and two levels of fully connected layers.

Furthermore, the number of the convolution kernels and the number of the bias matrixes of the convolution layer of each level and the numbers of the weight matrixes and bias vectors of each level of fully connected layer are not specifically limited in the embodiment of the present disclosure. In addition, dimensions of each convolution kernel and bias matrix and dimensions of each weight matrix and each bias vector are also not limited in the embodiment of the present disclosure. During specific implementation, the numbers and dimensions of the convolution kernels and bias matrixes of the convolution layer of each level and the numbers and dimensions of the weight matrixes and bias vectors of each level of fully connected layer may all adopt empirical values.

For example, the convolution layer of the first level may include 96 convolution kernels C¹with sizes of 5×5, C¹={C_i¹}_i=1⁹⁶, and 96 bias matrixes B¹with sizes of 5×5, B¹={B_i¹}_i=1⁹⁶.

With reference to the abovementioned content, when the initial model parameters of the CNN model to be trained are acquired, a value may be randomly selected from a specified numerical value range as a value of each of the elements in the initial model parameters. For example, for each of the elements in each initial convolution kernel, initial weight matrix, initial bias matrix and initial bias vector, a random number may be selected from interval [−r, r]. Here, r is a threshold value of the initial model parameter, and it may be an empirical value. For example, r may be 0.001.

In step 202, a plurality of training images are acquired.

The classifications of the training images are related to the type of the recognition of the CNN model to be trained. For example, when the CNN model to be trained is a model capable of recognizing different languages, the training images may be images including characters or character strings in different languages. When the CNN model to be trained is a model which recognizes colors, the training images are images including different colors and the like.

Optionally, with continuous development of a global economic situation, transnational communication and cooperation frequencies of nations, enterprises and persons greatly increase. In such case, a person may usually encounter various places where multilingual information is required to be used or understood inevitably, such as ordering at an exotic restaurant, participation in an international conference and browsing of a foreign website. This means that a technology for processing and understanding multilingual information has become a very important and urgent need. Specifically, such a technology requires automatic recognition of a language classification of a word obtained from a complicated natural scene. The recognition of a language classification of a word in a natural scene is a very tough challenge. There exist some major difficulties. One difficulty lies in that a font, color and layout manner of word information in the natural scene are usually messy, and its background is more complicated and may include more noise information. Another difficulty lies in that alphabets of different languages may include some characters which are completely the same or highly similar, and for example, there are some characters with the same shapes in English and Russian. Thus, it is difficult to recognize a language classification only from appearances of words. The last difficulty lies in that many classifiers for pictures with fixed width-height ratios become inapplicable to word pictures with completely inconsistent width-height ratios in the natural scene.

With reference to the abovementioned content, a CNN model capable of recognizing an image including characters in different languages in a natural scene is disclosed, and may implement rapid recognition of the language classifications of the characters in the image in the natural scene. Therefore, in the embodiment of the present disclosure, the training images may be images in the natural scene, and the images in the natural scene include characters in different languages. On such a basis, the CNN model to be trained may be a language recognition classifier.

In addition, sizes of the training images are not specifically limited in the embodiment of the present disclosure, and may be determined with reference to classifications of the training images during specific implementation. For example, an image including characters in different languages in the natural scene is usually greatly different in a horizontal direction and relatively similar in a vertical direction. That is, features of such an image are usually reflected in the horizontal direction. Therefore, in the embodiment of the present disclosure, for each of the training images, a specific value of its height may be limited, while the value of its width is not limited.

Specifically, in the embodiment of the present disclosure, all the training images may have the same height and width, that is, all the training images have a specified height and a specified width. In addition, all the training images may have the same height but different widths. For example, all the training images have the specified height but different widths. The specified height is a first number of pixels, and the specified width is a second number of pixels. Specifically, the first number may be an empirical value, and for example, the first number may be 32. The second number may be a random number selected within a preset range. In addition, a value of the second number may be related to classification of the image. For example, characters of the image including the characters in different languages in the natural scene are usually distributed in the horizontal direction, so that its width is usually larger than the height, and on such a basis, the first number is usually smaller than the second number. From the content of this part, the preset range may be from 48 to 256.

With reference to the size classifications of the training images, when the plurality of training images are acquired, the following two manners are adopted without limitation.

In the first manner, all the training images have a specified height and a specified width.

In such a manner, a plurality of initial training images may be acquired at first, and for each of the plurality of initial training images, a width-height ratio of the initial training image may be kept, and the initial training image is processed into a first image with the specified height; and then, the first image is processed into a second image with the specified width. At this moment, a training image which corresponds to the initial training image and has the specified height and the specified width may be obtained. All of the initial training images are processed in the same manner, and then the plurality of training images may be obtained.

For example, for a certain initial training image with a size of 64 pixels*120 pixels, when the specified height includes 32 pixels, the initial training image may be processed to obtain a training image with a size of 32 pixels*60 pixels.

In such a manner, when the first image is processed to obtain the second image with the specified width, since width of the first image may be larger than the specified width, and may also be smaller than the specified width. For the two cases, there may exist the following two processing manner for obtaining the image with the specified width.

In the first case where the width of the first image is smaller than the specified width, left and right sides of the first image are uniformly filled with pixels having a specified gray-scale value until the width of the first image reaches the specified width.

A specific numerical value adopted as the specified gray-scale value may be set according to a requirement or by experiences. For example, the specified gray-scale value may be 0.

For example, when the specified width is 120 pixels and the width of a certain first image is 116 pixels, the left side of the first image may be filled with 2 pixels with the specified gray-scale value, and the right side of the first image may be filled with 2 pixels with the specified gray-scale value.

In the second case where the width of the first image is larger than the specified width, pixels on the left and right sides of the first image may be uniformly cropped until the width of the first image reaches the specified width.

For example, when the specified width is 120 pixels and the width of a certain first image is 124 pixels, 2 pixels may be cropped from the left side of the first image, and 2 pixels may be cropped from the right side of the first image.

When the CNN model is trained by using the plurality of training images obtained in such a manner, all the training images may be ensured to have the same height and width. In such case, when the CNN model to be trained is trained, a plurality of processes may be set in a computer, and each process may be controlled to process a training image, so that the speed at which the CNN model is trained may be increased.

In the second manner, all the training images have the specified height, but the widths of each of the training images are not limited.

In such a manner, the plurality of initial training images may be acquired at first, and for each of the initial training images in the plurality of initial training images, a width-height ratio of the initial training image may be kept, the initial training image is processed to obtain an image with the specified height, and a width corresponding to the specified height is determined as a width of the initial training image. All of the initial training images are processed in the same way, and then the plurality of training images may be obtained.

For example, for a certain initial training image, when it is processed to obtain an image with the specified height, the width corresponding to the specified height is 120 pixels, then the height of the training image obtained in such a manner is the specified height, and the width is 120 pixels.

When the CNN model is trained by using the plurality of training images obtained in such a manner, the training images have the same specified height but different widths. At this moment, when the CNN model to be trained is trained, the trained CNN model may be ensured to have relatively higher recognition accuracy.

It is to be noted that the plurality of training images obtained in the first manner may be used when the CNN model to be trained is specifically trained by using the plurality of acquired training images, thereby implementing rapid acquisition of the trained CNN model, or the plurality of training images obtained in the second manner may be used, thereby ensuring that the trained CNN model has relatively higher recognition accuracy. In addition, the CNN model may also be trained by using the training images obtained in both the first manner and the second manner, thereby increasing the training speed of the CNN model and simultaneously ensuring that the trained CNN model has relatively higher recognition accuracy.

In step 203, on the convolution layer of each level, convolution operation and maximal pooling operation are performed on each of the training images to obtain a first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix of the convolution layer of each level.

By the convolution operation and the maximal pooling operation, features of different levels of the images may be extracted more and more deeply. Therefore, when the CNN model to be trained is trained, the convolution operation and the maximal pooling operation may be performed on the training images on the convolution layer of each level to obtain the features of the training images.

When the convolution operation is performed, it is necessary to use the convolution kernels and bias matrixes of the convolution layer of each level. When the maximal pooling operation is performed, reference will be made to the convolutional images obtained by convolution. One convolutional image may be obtained by the convolution operation on each of the different convolution layers and one feature image may be obtained by the maximal pooling operation on each convolutional image. Thus, images obtained by the convolution operation on the convolution layer of each level are called the convolutional images on the convolution layer of each level, and images obtained by the maximal pooling operation on the convolutional images on the convolution layer of each level are called the first feature images on the convolution layer of each level in the embodiment of the present disclosure.

With reference to the abovementioned content, the step that the convolution operation and the maximal pooling operation are performed on each of the training images to obtain the first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix on the convolution layer of each level includes, but not limited to, the following Step 2031 to Step 2033.

In Step 2031, for each of the training images, the first feature image on the convolution layer of a previous level is input to a current convolution layer, and the convolution operation is performed on the first feature image on the convolution layer of the previous level to obtain the convolutional image on the current convolution layer by using the initial convolution kernel and initial bias matrix of the current convolution layer. The first feature image on the convolution layer of the previous level is the training image if the current convolution layer is the convolution layer of the first level.

Specifically, for each of the training images, if the current convolution layer is the convolution layer of the first level, the training image may be transmitted to the convolution layer of the first level at first, and the convolution operation is performed on the training image to obtain the convolutional image on the convolution layer of the first level by using the initial convolution kernel and initial bias matrix on the convolution layer of the first level. Furthermore, the maximal pooling operation is performed on the convolutional image on the convolution layer of the first level to obtain the first feature image on the convolution layer of the first level. The first feature image on the convolution layer of the first level is transmitted to the convolution layer of the second level, and the convolution operation and the maximal pooling operation are performed on the convolution layer of the second level.

When the convolution operation is performed on the first feature image on the convolution layer of the previous level by using the initial convolution kernel and initial bias matrix on the current convolution layer, the convolution operation may be performed on the first feature image on the convolution layer of the previous level by using each convolution kernel in the current convolution layer, and a sliding step length of each convolution kernel in the current convolution layer on the first feature image on the convolution layer of the previous level is 1.

Specifically, if the current convolution layer is, for example, the convolution layer of the first level, when the convolution operation is performed on the training image by using a certain convolution kernel in the convolution layer of the first level, the following formula may be adopted:

D_i¹=I*C_i¹+B_i¹

where D_i¹represents the convolutional image, obtained by performing the convolution operation on the training image I by using the i^thconvolution kernel C_i¹in the convolution kernels on the convolution layer of the first level, on the convolution layer of the first level, and B_i¹represents the i^thbias matrix in the bias matrixes B¹on the convolution layer of the first level.

Optionally, considering a property of the convolution operation, when the convolution operation is performed, the convolution operation might not be performed on the pixels on an edge of the image, so that the size of the convolutional image obtained by the convolution operation is changed. In order to avoid such a condition, before the convolution operation is performed on the convolution layer of each level, pixel filling may also be performed on the image on which the convolution operation is to be performed, and then a size of the image on which the convolution operation is to be performed may be kept unchanged after the convolution operation is performed.

Specifically, when the image on which the convolution operation is to be performed is filled, a periphery of the image on which the convolution operation is to be performed may be filled by using a preset gray-scale value, such as “0”. The interval at which the filling is performed is related to the dimension of the convolution kernel. When the dimension of the convolution kernel is n, the step length at which the filling is performed is (n−1)/2. For example, when the convolution kernel is 3*3, the step length at which the filling is performed is 1 pixel.

The image on which the convolution operation is to be performed in the step is an image input to the convolution layer of each level. For example, when the current convolution layer is the convolution layer of the first level, the image on which the convolution operation is to be performed is the training image. When the current convolution layer is not the convolution layer of the first level, the image on which the convolution operation is to be performed is the first feature image on the convolution layer of the previous level.

In Step 2032, after the maximal pooling operation is performed on the convolution image on the current convolution layer to obtain the first feature image of the training image on the current convolution layer. The first feature image on the current convolution layer is then transmitted to the convolution layer of the next level.

Specifically, when the maximal pooling operation is performed on the convolutional image on the current convolution layer, a sliding box with a specified size, such as a 3×3 sliding box, may be adopted to slide from left to right and from top to bottom in each of the convolutional images on the current convolution layer, the sliding step length is a preset number of pixels, and for example, the preset number of pixels are 2 pixels. When the sliding box slides on each convolutional image, the elements, which are smaller than 0, of all elements contained in the sliding box are set to be 0. Then, a maximum element of all the elements is extracted, and a new image is reconstructed according to a sliding sequence. After the operation is finished, the first feature image on the current convolution layer is obtained.

The length and width of the first feature image on the current convolution layer may be calculated according to a property of the maximal pooling operation, and may be represented by the following formula:

w
=

floor


(

w
0

-
m

1

)

+
1

where floor represents a rounding-down function. For example, floor(2.5)=2. w represents the length or width of the first feature image on the current convolution layer, in is the number of pixels contained in a length or width of the sliding box, l is the sliding step length, and w₀is the height or with of the convolutional image on the current convolution layer. For example, when the size of the convolutional image on the current convolution layer is 32×w₀, the sliding box is 3*3 and the sliding step length is 2, the size of the first feature image, obtained by the maximal pooling operation, on the current convolution layer is 15×w₁,

w
1

=

floor


(

w
0

-
3

2

)

+
1.

In Step 2033, the convolution operation and the maximal pooling operation are performed on the first feature image on the current convolution layer on the convolution layer of the next level until the convolution operation and the maximal pooling operation are performed on the convolution layer of a last level to obtain the first feature image on the convolution layer of the last level.

A principle for performing the convolution operation and the maximal pooling operation on the convolution layer of each level is consistent with principles in Step 2031 and Step 2032. For detail. please refer to the contents in Step 2031 and Step 2032. and the detailed description will not be elaborated herein.

In Step 204, horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain a second feature image of each of the training images on the convolution layer of each level.

When the horizontal pooling operation is performed, the horizontal pooling operation may be performed on the first feature images on all the convolution layers, and the horizontal pooling operation may also be performed on the first feature images on some of the convolution layers. For example, if the CNN model to be trained includes four levels of convolution layers, the horizontal pooling operation may be performed on the convolution layer of each level, or the horizontal pooling operation may be performed on only the convolution layer of the second level, the convolution layer of the third level and the convolution layer of the fourth level.

Specifically, the step that the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain the second feature image of each of the training images on the convolution layer of each level includes, but not limited to, the following Step 2041 to Step 2043.

In Step 2041, for the first feature image of each training image on the convolution layer of each level, a maximum value of elements of each of rows of each of the images in the first feature image on the convolution layer are extracted. The first feature image includes a preset number of images, and the preset number is the same as each of the numbers of the convolution kernels and bias matrixes of the convolution layer.

For example, when the horizontal pooling operation is performed on the first feature image on the convolution layer of the second level, if the first feature image on the convolution layer of the second level is R², a process of performing the horizontal pooling operation on R²to obtain the second feature image H¹on the convolution layer of the second level is as follows:

if the ith feature image R_i²in R²is an image of which a height and width are 7 and w₂respectively, a maximum value of all elements in each of rows of R_i²is extracted.

In Step 2042, the maximum values extracted from all the rows of each image are arranged into a one-dimensional vector according to arrangement of pixels of each image.

In combination with the example in Step 2041, the maximum values of respective rows are arranged into a one-dimensional vector with a length of 7 according to top-down arrangement of pixels of each image. FIG. 4 shows a schematic diagram illustrating a process of performing horizontal pooling operation on a certain image in a first feature image on the convolution layer of a second level.

In Step 2043, the one-dimensional vectors of all the images in the first feature image on the convolution layer are combined to obtain the second feature image on the convolution layer.

In combination with the example in Step 2041, if R²includes 256 images, after the abovementioned operation is repeatedly performed on all of the 256 images in R², 256 one-dimensional vectors with the length of 7 are obtained. Each of one-dimensional vectors is considered as a column vector, and the one-dimensional vectors are sequentially combined into an image H′ of which a height and width are 7 and 256 respectively. H′ is the second feature image, obtained by the horizontal pooling operation over R², on the convolution layer of the second level.

In combination with the CNN model to be trained shown in FIG. 3, the CNN model to be trained includes four levels of convolution layers and two levels of fully connected layers, and the convolution layers of the respective levels include the same or different numbers of convolution kernels and bias matrixes. On such a basis, when the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain the second feature image of each of the training images on the convolution layer of each level, the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of the second level, the first feature image of each of the training images on the convolution layer of the third level and the first feature image of each of the training images on the convolution layer of the fourth level to obtain the second feature image of each of the training images on the convolution layer of the second level, the second feature image of each of the training images on the convolution layer of the third level and the second feature image of each of the training images on the convolution layer of the fourth level, respectively respectively.

In Step 205, a feature vector of each of the training images are determined according to the second feature image of each of the training images on the convolution layer of each level.

Specifically, for each of the training images, its feature vector is obtained by cascading the second feature images of the training image on the convolution layer of each level and connecting the elements in all the rows in the second feature images on the convolution layer of each level head to tail.

In combination with the CNN model to be trained shown in FIG. 3, for each of the training images, the feature vector of the training image may be determined according to the second feature image of the training image on the convolution layer of the second level, the second feature image of the training image on the convolution layer of the third level and the second feature image of the training image on the convolution layer of the fourth level.

Specifically, for a certain training image, the second feature image on the convolution layer of the second level, the second feature image of the training image on the convolution layer of the third level and the second feature image of the training image on the convolution layer of the fourth level may be cascaded, and the elements in all the rows in the second feature image on the convolution layer of the second level, the second feature image of the training image on the convolution layer of the third level and the second feature image of the training image on the convolution layer of the fourth level are connected head to tail to obtain the feature vector of the training image.

For example, if the second feature image of the training image on the convolution layer of the second level is H¹, the second feature on the convolution layer of the third level is H²and the second feature image of the training image on the convolution layer of the fourth level is H³, elements in all rows in H¹may be connected head to tail to obtain a first one-dimensional vector; elements in all rows in H²are connected head to tail to obtain a second one-dimensional vector; and elements in all rows in H³are connected head to tail to obtain a third one-dimensional vector. The first one-dimensional vector, the second one-dimensional vector and the third one-dimensional vector are further connected head to tail to obtain the feature vector of the training image.

In Step 206, each feature vector is processed to obtain a classification probability vector of each of the training images according to the initial weight matrixes and the initial bias vectors.

Specifically, a feature vector of a training image may be input to the fully connected layer, the feature vector is multiplied by the weight matrix of the fully connected layer, and a product result is added to the bias vector of the fully connected layer, thereby obtain a classification probability vector of the training image.

The classification probability vector is a one-dimensional vector, and the number of elements in the classification probability vector is related to the number of image classifications which may be recognized by the CNN model to be trained. For example, if 10 classifications of images may be recognized by using the training images, the classification probability vector includes 10 elements. Each of the elements represents a probability that the training image belongs to a certain classification.

In Step 207, a classification error is calculated according to the classification probability vector and initial classification of each of the training images.

The step that the classification error is calculated according to the classification probability vector and initial classification of each of the training images includes, but not limited to, the following Step 2071 to Step 2073.

In Step 2071, initial classifications of each of the training images are acquired.

For each of the training images, its initial classification is known. The initial classification may be manually extracted and recorded. On such a basis, the initial classification of each of the training images may be acquired directly according to recorded data.

In Step 2072, the classification error of each of the training images is calculated according to the classification probability vector and initial classification of each of the training images by using the following formula:

Loss=−Lny_label,

where Loss represents the classification error of each of the training images, label represents the initial classification of each of the training images, y_irepresents a element of the classification probability vector of each of the training images, and y_labelrepresents a classification probability corresponding to the initial classification. For example, when the initial training image is of a first classification, y_labelis y₁, i.e. a first element in the classification probability vector.

In Step 2073, the mean of the classification errors of all the training images is calculated, and the mean of the classification errors is determined as a classification error.

Of course, the process of calculating the classification errors is described by taking the mean of the classification errors of all the training images as an example. During specific implementation, when the classification errors are calculated, another manner may also be adopted, and will not be specifically limited in the embodiment of the present disclosure.

In Step 208, the model parameters of the CNN model to be trained are regulated on the basis of the classification errors.

Specifically, when the model parameters of the CNN model to be trained is regulated, a Stochastic Gradient Descent (SGD) algorithm is adopted without limitation to reversely transmit the classification errors back to the CNN model to be trained, thereby updating all the elements in the model parameters of the CNN model to be trained.

A specific implementation process of reversely transmitting the classification errors back to the CNN model to be trained to update all the elements in the model parameters of the CNN model to be trained by using the SGD algorithm will not be specifically limited in the embodiment of the present disclosure.

In Step 209, the process of determining the classification errors is continued on the basis of the regulated model parameters and the plurality of training images until the number of iterations reaches a preset number.

Specifically, Step 203 to Step 208 are repeated by using the plurality of training images acquired in Step 202 and the regulated model parameters obtained by Step 208 until the number of repetition times, i.e. the number of iterations, reaches the preset number.

A specific range of the numerical value of the preset number is not specifically limited in the embodiment of the present disclosure. In order to make the CNN model to be trained relatively more accurate, a value of the preset number may be relatively larger. During specific implementation, the preset number may be an empirical value, and for example, is 300,000.

In Step 210, model parameters obtained when the number of iterations reaches the preset number are determined as the model parameters of the trained CNN model.

For example, if the preset number is 300,000, the model parameters obtained when the number of iterations reaches 300,000 are determined as the model parameters of the trained CNN model.

In Step 211, an image to be recognized is input into the trained CNN model, a classification probability vector of the image to be recognized is calculated, and a classification corresponding to an element, which is the most approximate to 1, of the elements of the classification probability vector is determined as the classification of the image to be recognized.

Specifically, if the classification probability vector includes 10 elements, each element corresponding to a classification, and the second element in the classification probability vector is the most approximate to 1, the image to be recognized belongs to a second classification.

According to the method provided by the embodiment of the present disclosure, after the convolution operation and the maximal pooling operation are performed on the training images on the convolution layer of each level, the horizontal pooling operation is further performed on the feature images obtained by the maximal cooling operation. During the horizontal cooling operation, the feature images which identify features of the feature images in a horizontal direction may further be extracted from the feature images to ensure that the trained CNN model may break limits of a size of an image to be recognized in the horizontal direction during image classification identification and recognize an image to be recognized with any size, so that an image recognition application range of the CNN model trained in such a manner is relatively wider.

With reference to the contents of the embodiment corresponding to FIG. 1 or FIG. 2 and the schematic diagram illustrating the CNN model to be trained shown in FIG. 3, a method provided by the embodiment of the present disclosure is explained with the condition that the CNN model to be trained includes four levels of convolution layers and two levels of fully connected layers as an example in the embodiment of the present disclosure. In addition, in the embodiment of the present disclosure, descriptions are made with horizontal pooling operation over first feature images on the convolution layer of a second level, first feature images on the convolution layer of the third level and first feature images on the convolution layer of the fourth level as an example in the embodiment of the present disclosure. As shown in FIGS. 5A-5B, a flow of the method provided by the embodiment of the present disclosure includes the following steps.

In Step 501, initial model parameters of a CNN model to be trained are acquired, wherein the initial model parameters include an initial convolution kernel and initial bias matrix of the convolution layer of the first level, an initial convolution kernel and initial bias matrix of the convolution layer of the second level, an initial convolution kernel and initial bias matrix of the convolution layer of the third level, an initial convolution kernel and initial bias matrix of the convolution layer of the fourth level, an initial weight matrix and initial bias vector of the fully connected layer of the first level and an initial weight matrix and initial bias vector of the fully connected layer of the second level.

The convolution layers of the respective levels may include the same or different numbers of convolution kernels and bias matrixes, and the fully connected layer of the respective levels include the same or different numbers of weight matrixes and bias vectors. The numbers of the convolution kernels and bias matrixes of the convolution layer of each level and the numbers of the weight matrixes and bias vectors of each level of fully connected layers are not specifically limited in the embodiment of the present disclosure. In addition, dimensions of each convolution kernel and bias matrix and dimensions of each weight matrix and each bias vector are also not limited in the embodiment of the present disclosure. During specific implementation, the numbers and dimensions of the convolution kernels and bias matrixes of the convolution layer of each level and the numbers and dimensions of the weight matrixes and bias vectors of each level of fully connected layer may adopt empirical values.

In addition, in the embodiment of the present disclosure, the convolution kernels, bias matrixes, weight matrixes and bias vectors in the convolution layer of each level in the initial model parameters are called initial convolution kernels, initial bias matrixes, initial weight matrixes and initial bias vectors. Therefore, when the CNN model to be trained is trained, the initial convolution kernels of the convolution layer of each level, the initial bias matrixes of the convolution layer of each level, the initial weight matrixes of the fully connected layers and the initial bias vectors of the fully connected layers will be acquired.

In order to facilitate explanation about the embodiment of the present disclosure, in the embodiment of the present disclosure, the method provided by the embodiment of the present disclosure is explained with the condition that the convolution layer of the first level includes 96 convolution kernels C¹with sizes of 5×5, C¹={C_i¹}_i=1⁹⁶, and, 96 bias matrixes B¹with sizes of 5×5, B¹={B_i¹}_i=1⁹⁶, the convolution layer of the second level includes 256 convolution kernels C²with sizes of 5×5, C²={C_i²}_i=1²⁵⁶, and 256 bias matrixes B²with sizes of 5×5, B²={B_i²}_i=1²⁵⁶, the convolution layer of the third level includes 384 convolution kernels C³with sizes of 3×3, C³={C_i³}_i=1³⁸⁴, and 384 bias matrixes B³with sizes of 3×3, B³={B_i³}_i=1³⁸⁴, the convolution layer of the fourth level includes 512 convolution kernels C⁴with sizes of 3×3, C⁴={C_i⁴}_i=1⁵¹², and 512 bias matrixes B⁴with sizes of 3×3, B⁴={B_i⁴}_i=1⁵¹², the fully connected layer of the first level includes a weight matrix W¹with a size of 3,456×1,024 and an bias vector B⁵with a length of 1,024, and the fully connected layer of the second level includes a weight matrix W²with a size of 1,024×10 and an bias vector B⁶with a length of 10 as an example.

When the initial model parameters of the CNN model to be trained are acquired, a value may be randomly selected from a specified numerical value range as a value of each element in the initial model parameter. For example, for each element in each initial convolution kernel, initial weight matrix, initial bias matrix and initial bias vector, a random number may be selected from interval [−r,r]. Here, r is a threshold value of the initial model parameter, and it may be an empirical value. For example, r may be 0.001.

In Step 502, a training image I is acquired, wherein the training image I has a specified height.

The training image I is one of a plurality of training images. In order to facilitate description, a process of training the CNN model to be trained is explained with processing of the training image I as an example in subsequent implementation of the present disclosure.

In addition, the training image I has the specified height and a specified width. Specifically, the specified height may be a first number of pixels, and the specified width may be a second number of pixels. Specifically, the first number may be an empirical value, and for example, the first number may be 32. The second number may be a random number selected within a preset range, wherein the preset range may be 48 to 256. In order to facilitate description, a size of the training image is, for example, 32*w₀in the subsequent embodiment of the present disclosure, wherein, 32 is the first number, and w₀is the second number.

In Step 503, on the convolution layer of the first level, convolution operation and maximal pooling operation are performed on the training image to obtain a first feature image R¹of the training image I on the convolution layer of the first level by using the convolution kernel and bias matrix of the convolution layer of the first level.

The step that the convolution operation and the maximal pooling operation are performed on the training image I on the convolution layer of the first level by using the initial convolution kernel and initial bias matrix of the convolution layer of the first level will be specifically described below with reference to each initial model parameters in Step 502. Specifically, the process of obtaining the first feature image R¹of the training image I on the convolution layer of the first level is implemented as follows.

In Step 5031, the training image I is filled to obtain a filled training image I_tr.

The step is an optional step. By the step, a convolutional image on the convolution layer of the first level obtained by the convolution operation on the convolution layer of the first level may be ensured to have the same size with the training image I. The process of filling has been described in Step 203, may specifically refer to the content in Step 203, and will not be elaborated herein.

For example, the size of the training image I is 32×w₀, and a periphery of the training image I may be uniformly filled with element “0”, wherein a filling step length is 2, so that a training image I_trwith a height and width of 36 pixels and w₀+4 pixels respectively is obtained.

In Step 5032, on the convolution layer of the first level, the convolution operation is performed on the filled training image I_trto obtain the convolutional image D¹on the convolution layer of the first level by using the convolution kernels and bias matrixes on the convolution layer of the first level.

Specifically, in combination with the example in Step 502, the convolution operation may be performed on the filled training image I_trby using each convolution kernel in C¹, and a sliding step length of each convolution kernel on the filled training image I_tris 1, that is:

D_i¹=I_tr*C_i¹+B_i¹

where D_i¹represents a convolutional image on the convolution layer of the first level obtained by performing the convolution operation on the filled training image I_trby using the ith convolution kernel C_i¹in the convolution layer of the first level, and B_i¹represents the ith bias matrix in the bias matrixes B¹in the convolution layer of the first level. According to a property of the convolution operation over an image, the convolutional image D¹={D_i¹}_i=1⁹⁶, on the convolution layer of the first level may be obtained, where the size of D_i¹is 32×w₀.

In Step 5033, the maximal pooling operation is performed on the convolutional image D¹on the convolution layer of the first level to obtain the first feature image R¹on the convolution layer of the first level.

A specific content about the maximal pooling operation has been elaborated in Step 203, may specifically refer to the content in Step 203, and will not be elaborated herein.

For example, in combination with the example in Step 502, a 3×3 sliding box may be adopted to slide from left to right and from top to bottom in each of the convolutional images D¹on the convolution layer of the first level, and the sliding step length is 2. When the sliding box slides on each convolutional image, elements, which is smaller than 0, of 9 elements in the box are set to be 0, then a maximum element in the 9 elements is extracted, and a new image is reconstructed according to the sliding sequence. After the operation is finished, the first feature image R¹on the convolution layer of the first level is obtained, R¹={R_i¹}_i=1⁹⁶. According to a property of the maximal pooling operation, a size of R¹is 15×w₁, where

w
1

=

floor


(

w
0

-
3

2

)

+
1

,

and floor represents a rounding-down function. For example, floor(2.5)=2. In addition, “3” in the numerator in the formula represents a size of a dimension of the sliding box, and “2” in the denominator represents the sliding step length.

It is to be noted that 3×3 here is only an example, and during specific implementation, the size of the sliding box may also be of another numerical value. The size of the sliding box will not be limited in the embodiment of the present disclosure.

In Step 504, on the convolution layer of the second level, the convolution operation and the maximal pooling operation are performed on the first feature image R¹on the convolution layer of the first level to obtain a first feature image R²on the convolution layer of the second level by using the convolution kernel and bias matrix of the convolution layer of the second level.

The step that the convolution operation and the maximal pooling operation are performed on the first feature image R¹on the convolution layer of the first level by using the convolution kernel and bias matrix of the convolution layer of the second level on the convolution layer of the second level will be specifically described below with reference to each initial model parameter. Specifically, the process of obtaining the first feature image R²on the convolution layer of the second level is implemented as follows.

In Step 5041, the first feature image R¹on the convolution layer of the first level is filled to obtain a filled training image Rp¹.

The principle of the step is consistent with the principle of Step 5031, and may specifically refer to the content in Step 5031. The step is an optional step. By the step, the image obtained by the convolution operation may be ensured to have the same size with the first feature image R¹on the convolution layer of the first level.

For example, in combination with the example in Step 502, a filling length includes, but not limited to, 2, the filled image Rp¹is Rp¹={Rp_i¹}_i=1⁹⁶, and a height and width of Rp_i¹, are 19 and w₁+4 respectively.

In Step 5042, on the convolution layer of the second level, the convolution operation is performed on the filled image Rp¹to obtain the convolutional image D²on the convolution layer of the second level by using the convolution kernels and bias matrixes on the convolution layer of the second level.

The principle of the step is consistent with the principle of step S032, may specifically refer to the content in Step 5032, and will not be elaborated herein.

Specifically, in combination with the example in the abovementioned step, the convolution operation may be performed on each image in Rp¹by using each convolution kernel in C², a sliding step length of each convolution kernel on the image is 1, and convolution results are added to obtain each convolutional image D_i²on the convolution layer of the second level. That is:

D
i
2

=

∑

j
=
1

96



Rp
j
1

*

C
i
2

+

B
i
2

where D_i²represents the ith obtained convolutional image on the convolution layer of the second level, C_i²represents the ith convolution kernel in C², B_i²represents the ith bias matrix in B², and Rp_j¹represents the jth image in Rp¹. According to the property of the convolution operation, D²={D_i²}_i=1²⁵⁶, where D_i²represents the convolutional image obtained by performing the convolution operation on Rp¹by using the ith convolution kernel C_i², and the size of D_i²is 15×w₁.

In Step 5043, the maximal pooling operation is performed on the convolutional image D²on the convolution layer of the second level to obtain the first feature image R²on the convolution layer of the second level.

The principle of the step is consistent with the step of Step 5033, and may specifically refer to the content in Step 5033. Specifically, R²={R_i²}_i=1²⁵⁶, the size of R_i²is 7×w₂, and

w
2

=

floor


(

w
1

-
3

2

)

+
1.

In Step 5044, horizontal pooling operation is performed on the first feature image R²on the convolution layer of the second level to obtain a second feature image H¹on the convolution layer of the second level.

Specifically, the ith feature image R_i²in R²is an image of which a height and width are 7 and w₂respectively, and the maximum value of all elements in each of rows of R_i²is extracted, and the maximum values are sequentially connected into a one-dimensional vector with a length of 7.

After the operation is repeatedly performed on all of the 256 feature images in R², 256 one-dimensional vectors with the length of 7 are obtained. The respective one-dimensional vectors are considered as column vectors, and are sequentially combined into an image H¹of which a height and width are 7 and 256 respectively. H¹is the second feature image, obtained by the horizontal pooling operation over R², on the convolution layer of the second level.

In Step 505, on the convolution layer of the third level, the convolution operation and the maximal pooling operation are performed on the first feature image R²on the convolution layer of the second level to obtain a first feature image R³on the convolution layer of the third level by using the convolution kernel and bias matrix of the convolution layer of the third level.

The step that the convolution operation and the maximal pooling operation are performed on the first feature image R²on the convolution layer of the second level by using the convolution kernel and bias matrix of the convolution layer of the third level on the convolution layer of the third level will be specifically described below with reference to each model parameters in Step 502. Specifically, the process of obtaining the first feature image R³on the convolution layer of the third level is implemented as follows.

In Step 5051, the first feature image R²on the convolution layer of the second level is filled to obtain a filled training image Rp².

The principle of the step is consistent with the principle of Step 5031, and may specifically refer to the content in Step 5031. The step is an optional step. By the step, the image obtained by the convolution operation may be ensured to have the same size with the first feature image R²on the convolution layer of the second level.

Specifically, a filling length includes, but not limited to, 1, and the new image Rp²is obtained by perform filling, wherein Rp²={Rp_i²}_i=1²⁵⁶, and a height and width of are 9 and w₂+2 respectively.

In Step 5052, on the convolution layer of the third level, the convolution operation is performed on the filled image Rp²to obtain a convolutional image D³on the convolution layer of the third level by using the convolution kernels and bias matrixes on the convolution layer of the third level.

The principle of the step is consistent with the principle of step 5032, may specifically refer to the content in Step 5032, and will not be elaborated herein.

A sliding step length of each convolution kernel on the filled image Rp²is 1, that is:

D
i
3

=

∑

j
=
1

256



Rp
j
2

*

C
i
3

+

B
i
3

where D_i³represents the ith obtained convolutional image on the convolution layer of the third level, C_i³represents the ith convolution kernel in C³, B_i³represents the ith bias matrix in B³, and Rp_j²represents the jth image in Rp². According to the property of the convolution operation, D³={D_i³}_i=1³⁸⁴, where D_i³represents the convolutional image obtained by performing the convolution operation on Rp²by using the ith convolution kernel C_i³, and the size of D_i³is 7×w₂.

In Step 5053, the maximal pooling operation is performed on the convolutional image D³on the convolution layer of the third level to obtain the first feature image R³on the convolution layer of the third level.

The principle of the step is consistent with the step of Step 5033, and may specifically refer to the content in Step 5033. Specifically, R³={R_i³}_i=1³⁸⁴, and the size of R_i³is 3×w₃, where

w
3

=

floor


(

w
2

-
3

2

)

+
1.

In Step 5054, horizontal pooling operation is performed on the first feature image R³on the convolution layer of the third level to obtain a second feature image H²on the convolution layer of the third level.

The principle of the step is consistent with the step of Step 5044, and may specifically refer to the content of Step 5044.

Specifically, the horizontal pooling operation may be performed on the first feature image R³on the convolution layer of the third level to obtain an image H²of which a height and width are 3 and 384 respectively. H²is output of the horizontal pooling operation over R³.

In Step 506, on the convolution layer of the fourth level, the convolution operation and the maximal pooling operation are performed on the first feature image R³on the convolution layer of the third level to obtain a first feature image R⁴on the convolution layer of the fourth level by using the convolution kernel and bias matrix of the convolution layer of the fourth level.

The step that the convolution operation and the maximal pooling operation are performed on the first feature image R³on the convolution layer of the third level by using the convolution kernel and bias matrix of the convolution layer of the fourth level on the convolution layer of the fourth level will be specifically described below with reference to each model parameters in Step 502. Specifically, the process of obtaining the first feature image R⁴on the convolution layer of the fourth level is implemented as follows.

In Step 5061, on the convolution layer of the fourth level, the convolution operation is performed on the first feature image R³on the convolution layer of the third level to obtain a convolutional image D⁴on the convolution layer of the fourth level by using the convolution kernels and bias matrixes on the convolution layer of the fourth level.

A sliding step length of the convolution kernel on R³is 1, that is:

D
i
4

=

∑

j
=
1

384



R
j
3

*

C
i
4

+

B
i
4

where D_i⁴, represents the ith obtained convolutional image on the convolution layer of the fourth level, C_i⁴represents the ith convolution kernel in C⁴, B_i⁴represents the ith bias matrix in B⁴, and R_j³represents the jth image in R³. According to the property of the convolution operation, D⁴={D_i⁴}_i=1⁵¹², where D_i⁴represents the convolutional image obtained by performing the convolution operation on R³by using the ith convolution kernel C_i⁴, and the size of D_i⁴is 1×(w₃−2). Since the maximal pooling operation is not performed on the first feature image on the convolution layer of the fourth level, the convolutional image is the first feature image on the convolution layer of the fourth level, that is, R⁴=D⁴, R⁴={R_i⁴}_i=1⁵¹², and the size of R_i⁴is 1×(w₃−2).

In Step 5062, the horizontal pooling operation is performed on the first feature image R⁴on the convolution layer of the fourth level to obtain a second feature image H³on the convolution layer of the fourth level.

The principle of the step is consistent with the step of Step 5044, and may specifically refer to the content of Step 5044.

Specifically, the ith feature image R_i⁴in R⁴is a vector with a length of w₃−2, and a maximum value in all elements in R_i⁴is extracted. After the operation is repeatedly performed on all 512 feature images in R⁴, 512 numbers are obtained, and these numbers are sequentially connected into a vector H³with a length of 512. H³is the output of the horizontal pooling operation over R⁴.

In Step 507, a feature vector Fc¹of the training image I is determined according to the second feature image H¹of the training image on the convolution layer of the second level, the second feature image H²of the training image on the convolution layer of the third level and the second feature image H³of the training image on the convolution layer of the fourth level.

Specifically, the column vectors of H¹and H²may be cascaded by connecting the column vectors of H¹and H²head to tail, and then the cascaded result is further cascaded with the vector H³to obtain a one-dimensional vector Fc¹. The vector Fc¹is determined as the feature vector of the training image I.

In combination with the examples in each of the abovementioned steps, the size of H¹is 7×256, the size of H²is 3×384 and the length of H²is 512, so that a length of Fc¹is 3,456.

In Step 508, the feature vector Fc¹is determined as input of the fully connected layer of the first level, and Fc¹is processed to calculate output Fc²of the fully connected layer of the first level by using the weight matrix and bias vector of the fully connected layer of the first level.

With reference to the content of Step 502, the output Fc²of the fully connected layer of the first level may be calculated by using the weight matrix W¹and bias vector B⁵of the fully connected layer of the first level by using the following formula:

Fc²=Fc¹*W¹+B⁵.

With reference to the content of Step 508, a length of Fc²is 1,024. Results obtained in each of the abovementioned steps may refer to specific numerical values in FIG. 3. Of course, the numerical values in FIG. 3 are only examples, and are not intended to form limits to the embodiment of the present disclosure.

In Step 509, the output of the fully connected layer of the first level is input to the fully connected layer of the second level, a classification probability vector Y of the training image I is calculated by using the weight matrix and bias vector of the fully connected layer of the second level, and a classification error Loss is calculated according to the classification probability vector Y and initial classification label of the training image I.

The principle of the step has been specifically described in Step 206 and Step 207, may specifically refer to the contents in Step 206 and Step 207, and will not be elaborated herein.

Specifically, in combination with the example, the output Fc²of the fully connected layer of the first level may be determined as input of the fully connected layer of the second level. From the content of Step 502, the weight matrix of the fully connected layer of the second level is W², the bias vector is B⁶, and the classification probability vector Y of the training image I is calculated where Y=Fc²*W²⁺ B⁶, and Y=[Y₁, Y₂, ΛΛ Y_i, Λ Λ, Y_n], where Y_irepresents a probability that the input image I is of an ith classification, and n represents the number of image classifications which may be recognized by the CNN model to be trained. Descriptions are made only with the condition that the CNN model to be trained can recognize 10 image classifications as an example in FIG. 3.

Furthermore, the operation that the classification error Loss is calculated according to the classification probability vector Y and initial classification label of the training image I may be implemented by using the following formula:

Loss=−Lny_label.

For example, if the initial classification of the training image I is a first classification, then label=1.

In Step 510, Step 502 to Step 509 are repeatedly performed on each of the training images, a mean Loss of classification errors of all the training images is calculated, and the mean of the classification errors is determined as a classification error.

In Step 511, the model parameters of the CNN model to be trained are regulated on the basis of the classification errors.

The principle of the step has been specifically described in Step 208, may specifically refer to the content of Step 208, and will not be elaborated herein.

Specifically, when the classification error is reversely transmitted back to the CNN model to be trained by using an SGD algorithm, a learning rate in the SGD algorithm may adopt 0.01, and the mean of the classification errors Loss calculated in Step 510 is reversely transmitted back to the CNN model to be trained to update all the elements in the convolution kernels C¹, C², C³and C⁴, the bias matrixes B¹, B², B³and B⁴, the weight matrixes W¹and W²and the bias vectors B⁵and B⁶.

In Step 512, Step 503 to Step 511 are iteratively operated for a preset number N, and model parameters obtained when the number of iterations reaches the preset number N are determined as the model parameters of the trained CNN model.

A specific range of the numerial value of N is not specifically limited in the embodiment of the present disclosure. In order to make the trained CNN model relatively more accurate, a value of N may be relatively larger. During specific implementation, N may be an empirical value. For example, N is 300,000. The model parameters obtained when the number of iterations reaches 300,000 are determined as the model parameters of the trained CNN model.

An test shows that the CNN model trained by Step 501 to Step 512 has relatively higher accuracy when the CNN model to be trained includes the four levels of convolution layers and the horizontal pooling operation is performed on the convolution layer of the second level, the convolution layer of the third level and the convolution layer of the fourth level, so that the training accuracy of the CNN model may be improved.

According to the method provided by the embodiment of the present disclosure, after the convolution operation and the maximal pooling operation are performed on the training images on the convolution layer of each level, the horizontal pooling operation is further performed on the feature images obtained by the maximal cooling operation. During the horizontal cooling operation, the feature images which identify features of the feature images in a horizontal direction may further be extracted from the feature images to ensure that the trained CNN model may break limits of a size of an image to be recognized in the horizontal direction during image classification identification and recognize an image to be recognized with any size, so that an image recognition application range of the CNN model trained in such a manner is relatively wider.

FIG. 6 is a schematic diagram of a structure of a device for training a CNN model, according to another embodiment of the present disclosure. The device for training the CNN model may be configured to execute the method for training the CNN model provided by the embodiment corresponding to any one of FIG. 1, FIG. 2 and FIGS. 5A-5B. As shown in FIG. 6, the device includes the following modules.

A first acquisition module 601 is configured to acquire initial model parameters of a CNN model to be trained, wherein the initial model parameters includes initial convolution kernels and initial bias matrixes of convolution layers of respective levels, and an initial weight matrix and an initial bias vector of a fully connected layer.

A second acquisition module 602 is configured to acquire a plurality of training images.

A feature extraction module 603 is configured to, on the convolution layer of each level, perform convolution operation and maximal pooling operation on each of the training images to obtain a first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix of the convolution layer of each level.

A horizontal pooling module 604 is configured to perform horizontal pooling operation on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain a second feature image of each of the training images on the convolution layer of each level.

A first determination module 605 is configured to determine a feature vector of each of the training images according to the second feature image of each of the training images on the convolution layer of each level.

A processing module 606 is configured to process each feature vector to obtain a classification probability vector of each of the training images according to the initial weight matrixes and the initial bias vectors.

A calculation module 607 is configured to calculate a classification error according to the classification probability vector and initial classification of each of the training images.

A regulation module 608 is configured to regulate the model parameters of the CNN model to be trained on the basis of the classification errors.

An iteration module 609 is configured to continue the process of regulating the model parameters on the basis of the regulated model parameters and the plurality of training images until the number of iterations reaches a preset number.

A second determination module 610 is configured to determine model parameters obtained when the number of iterations reaches the preset number as the model parameters of the trained CNN model.

According to the device provided by the embodiment of the present disclosure, after the convolution operation and the maximal pooling operation are performed on the training images on the convolution layer of each level, the horizontal pooling operation is further performed on the feature images obtained by the maximal cooling operation. During the horizontal cooling operation, the feature images which identify features of the feature images in a horizontal direction may further be extracted from the feature images to ensure that the trained CNN model may break limits of a size of an image to be recognized in the horizontal direction during image classification identification and recognize an image to be recognized with any size, so that an image recognition application range of the CNN model trained in such a manner is relatively wider.

In another embodiment, the second acquisition module 602 includes the following units.

A first acquisition unit is configured to acquire a plurality of initial training images.

A first processing unit is configured to, for each of the initial training images, keep a width-height ratio of the initial training image, and process the initial training image to obtain a first image with a specified height.

A second processing unit is configured to process the first image to obtain a second image with a specified width.

A first determination unit is configured to determine the image with the specified height and the specified width as the training image corresponding to the initial training image.

In another embodiment, the second acquisition module 602 includes the following units.

A second acquisition unit is configured to acquire a plurality of initial training images.

A third processing unit is configured to, for each of the initial training images, keep a width-height ratio of the initial training image, process the initial training image to obtain an image with a specified height, and determine a width corresponding to the specified height as the width of the initial training image.

In another embodiment, the first processing unit includes the following subunits.

A processing subunit is configured to process each initial training image into the images with the specified height to obtain an initial width.

A filling subunit is configured to, when the initial width is smaller than the specified width, uniformly fill left and right sides of the processed initial training images with pixels having a specified gray-scale value until the initial width reaches the specified width.

A cropping subunit is configured to, when the initial width is larger than the specified width, uniformly crop pixels on the left and right sides of the processed initial training images until the initial width reaches the specified width.

In another embodiment, the feature extraction module 603 includes the following units.

A convolution unit is configured to, for each of the training images, input the first feature image on the convolution layer of a previous level to a current convolution layer, and perform the convolution operation on the first feature image on the convolution layer of the previous level to obtain a convolutional image on the current convolution layer by using the initial convolution kernel and initial bias matrix of the current convolution layer, wherein the first feature image on the convolution layer of the previous level is the training image if the current convolution layer is the convolution layer of the first level.

A maximal pooling unit is configured to perform the maximal pooling operation on the convolutional image on the current convolution layer to obtain the first feature image of the training image on the current convolution layer.

A transmission unit is configured to continue transmitting the first feature image on the current convolution layer to the convolution layer of a next level, and perform the convolution operation and the maximal pooling operation on the convolution layer of the next level until the convolution operation and the maximal pooling operation are performed on the convolution layer of a last level to obtain the first feature image on the convolution layer of the last level.

In another embodiment, the horizontal pooling module 604 includes the following units.

An extraction unit is configured to, for the first feature image of each training image on the convolution layer of each level, extract a maximum value of elements of each of rows of each of the images in the first feature image on the convolution layer, wherein the first feature image includes a preset number of images, and the preset number is the same as each of the numbers of the convolution kernels and bias matrixes of the convolution layer.

An arrangement unit is configured to arrange the maximum values extracted from all the rows of each image into a one-dimensional vector according to arrangement of pixels of each image.

A combination unit is configured to combine the one-dimensional vectors of all the images in the first feature image on the convolution layer to obtain the second feature image on the convolution layer.

In another embodiment, the first determination module 605 is configured to, for each of the training images, connect elements of all rows of the second feature image of the training image on the convolution layer of each level head to tail to obtain the feature vector of the training image.

In another embodiment, the calculation module 607 includes the following unit.

An acquisition unit is configured to acquire initial classifications of each of the training images.

A first calculation unit is configured to calculate the classification error of each of the training images according to classification probability vector and the initial classification of each of the training images by using the following formula:

Loss=−Lny_label,

where Loss represents the classification error of each of the training images, label represents the initial classification of each of the training images, y_irepresents a element of the classification probability vector of each of the training images, and y_labelrepresents a classification probability corresponding to the initial classification.

A second calculation unit is configured to calculate a mean of the classification errors of all the training images, and determine the mean of the classification errors as a classification error.

In another embodiment, the training images are images in a natural scene, the images in the natural scene include characters in different languages, and the CNN model to be trained is a language recognition classifier.

In another embodiment, the CNN model to be trained includes four levels of convolution layers and two fully connected layers, and the convolution layers of the respective levels include the same or different numbers of convolution kernels and bias matrixes.

In said another embodiments, the horizontal pooling module is configured to perform the horizontal pooling operation on the first feature image of each of the training images on the convolution layer of the second level, the first feature image of each of the training images on the convolution layer of the third level and the first feature image of each of the training images on the convolution layer of the fourth level to obtain the second feature image of each of the training images on the convolution layer of the second level, the second feature image of each of the training images on the convolution layer of the third level and the second feature image of each of the training images on the convolution layer of the fourth level, respectively.

In said another embodiments, the first determination module is configured to, for each of the training images, determine the feature vector of the training image according to the second feature image of the training image on the convolution layer of the second level, the second feature image of the training image on the convolution layer of the third level and the second feature image of the training image on the convolution layer of the fourth level.

All of the optional technical solutions may be freely combined into optional embodiments of the present disclosure, which will not be elaborated one by one.

FIG. 7 shows a schematic diagram of a structure of a terminal involved in an embodiment of the present disclosure, and the terminal may be configured to implement the method for training the CNN model provided by the embodiment corresponding to FIG. 1, FIG. 2 or FIG. 4. Details of the terminal is as follows.

The terminal 700 may include parts such as a Radio Frequency (RF) circuit 110, a memory 120 including one or more computer-readable storage medium, an input unit 130, a display unit 140, a sensor 150, an audio circuit 160, a Wireless Fidelity (WiFi) module 170, a processor 180 including one or more processing cores and a power supply 190. Those skilled in the art should know that the terminal structure shown in FIG. 7 is not intended to form limits to the terminal, and may include more or fewer parts than those shown in the figure, or some parts may be combined, or the parts may be arranged in different manners.

The RF circuit 110 may be configured to receive and send a signal in an information transmission and receiving or communication process, particularly transmit downlink information of a base station to one or more than one processor 180 for processing after receiving it, and also send involved uplink data to the base station. Usually, the RF circuit 110 includes, but not limited to, an antenna, at least one amplifier, a tuner, one or more oscillators, a Subscriber Identity Module (SIM) card, a transceiver, a coupler, a Low Noise Amplifier (LNA), a duplexer and the like. In addition, the RF circuit 110 may further communicate with a network and other equipment in a wireless communication manner. Wireless communication may adopt any communication standard or protocol, including, but not limited to, Global System of Mobile Communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), electronic mail, Short Messaging Service (SMS) and the like.

The memory 120 may be configured to store software programs and modules, and the processor 180 runs the software programs and modules stored in the memory 120, thereby executing various function applications and data processing. The memory 120 may mainly include a program storage area and a data storage area, wherein the program storage area may store operating systems, application programs required by at least one function (such as a sound play function and an image play function) and the like; and the data storage area may store data (such as audio data and a telephone book) created by use of the terminal 700 and the like. In addition, the memory 120 may include a high speed random access memory, and may also include a nonvolatile memory, such as at least one disk memory, a flash memory or another volatile solid state memory. Correspondingly, the memory 120 may further include a memory controller for providing access to the memory 120 for the processor 180 and the input unit 130.

The input unit 130 may be configured to receive input digital or character information and generate keyboard, mouse, operating rod, optical or trackball signal input related to user setting and function control. Specifically, the input unit 130 may include a touch-sensitive surface 131 and other input equipment 132. The touch-sensitive surface 131, also called a touch display screen or a touchpad, may collect touch operation of a user thereon or nearby (such as operation of the user executed on the touch-sensitive surface 131 or nearby the touch-sensitive surface 131 by using any proper object such as a finger and a stylus) and drive a corresponding connecting device according to a preset program. Optionally, the touch-sensitive surface 131 may include two parts, i.e. a touch detection device and a touch controller, wherein the touch detection device detects a touch direction of the user, detects a signal generated by the touch operation, and sends the signal to the touch controller; and the touch controller receives touch information from the touch detection device, converts it into a touch coordinate for sending to the processor 180, and may receive and execute a command sent by the processor 180. In addition, the touch-sensitive surface 131 may be implemented by adopting a resistor, a capacitor, infrared rays, surface acoustic waves and the like. Besides the touch-sensitive surface 131, the input unit 130 may further include the other input equipment 132. Specifically, the other input equipment 132 may include, but not limited to, one or more of a physical keyboard, a function key (such as a volume control key and an on/off key), a trackball, a mouse, an operating rod and the like.

The display unit 140 may be configured to display information input by the user or information provided for the user and various graphic user interfaces of the terminal 700, and these graphic user interfaces may be formed by graphics, texts, icons, videos and any combinations thereof. The display unit 140 may include a display panel 141, and the display panel 141 may be optionally configured in form of Liquid Crystal Display (LCD), Organic Light-Emitting Diode (OLED) and the like. Furthermore, the touch-sensitive surface 131 may cover the display panel 141, the touch operation on or nearby the touch-sensitive surface 131 is sent to the processor 180 to determine a classification of a touch event after being detected, and then the processor 180 provides corresponding visual output on the display panel 141 according to the classification of the touch event. Although the touch-sensitive surface 131 and display panel 141 in FIG. 7 realize input and output functions as two independent parts, the touch-sensitive surface 131 and the display panel 141 may be integrated to realize the input and output functions in some embodiments.

The terminal 700 may further include at least one sensor 150, such as a light sensor, a motion sensor and another sensor. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor may regulate luminance of the display panel 141 according to brightness of ambient light, and the proximity sensor may turn off the display panel 141 and/or backlight when the terminal 700 moves to an ear. As a motion sensor, a gravity acceleration sensor may detect a magnitude of acceleration in each direction (usually three axes), may detect a magnitude and direction of gravity under a still condition, and may be configured to identify a state application of the terminal (such as horizontal and vertical screen switching, a related game and magnetometer state calibration), a vibration identification related function (such as a pedometer and knocking) and the like; and the other sensor, such as a gyroscope, an air pressure gauge, a hygrometer, a thermometer and an infrared sensor, which may be configured in the terminal 700 will not be elaborated herein.

The audio circuit 160, a loudspeaker 161 and a microphone 162 may provide an audio interface between the user and the terminal 700. The audio circuit 160 may transmit an electric signal converted from received audio data to the loudspeaker 161, and the loudspeaker 161 converts the electric signal into a sound signal for output; and on the other hand, the microphone 162 converts a collected sound signal into an electric signal, the audio circuit 160 receives and converts the electric signal into audio data, outputs the audio data to the processor 180 processing, and sends the audio data to, for example, another terminal through the RF circuit 110, or outputs the audio data to the memory 120 for further processing. The audio circuit 160 may further include an earplug hole for providing communication between an external earphone and the terminal 700.

WiFi belongs to a short-distance wireless transmission technology, the terminal 700 may help the user to send and receive electronic mails, browse web pages, access streaming media and the like through the WiFi module 170, and it provides wireless wideband Internet access for the user. Although the WiFi module 170 is shown in FIG. 7, it may be understood that it is not a necessary component of the terminal 700 and may completely be eliminated without changing the scope of the nature of the present disclosure according to a requirement.

The processor 180, as a control center of the terminal 700, connects each part of the whole terminal by using various interfaces and lines, and executes various functions and data processing of the terminal 700 by running or executing the software programs and/or modules stored in the memory 120 and calling the data stored in the memory 120, thereby monitoring the whole terminal. Optionally, the processor 180 may include one or more processing cores; and preferably, the processor 180 may integrate an application processor and a modulation and demodulation processor, wherein the application processor mainly processes an operating system, a user interface, an application program and the like, and the modulation and demodulation processor mainly processes wireless communication. It may be understood that the modulation and demodulation processor may also not be integrated into the processor 180.

The terminal 700 further includes the power supply 190 (such as a battery) for supplying power to each part, and optionally, the power supply may be logically connected with the processor 180 through a power management system, thereby realizing functions of charging and discharging management, power consumption management and the like through the power management system. The power supply 190 may further include one or more than one direct current or alternating current power supply, rechargeable system, power failure detection circuit, power converter or inverter, power state indicator and the like.

Although not being shown, the terminal 700 may further include a camera, a Bluetooth module and the like, which will not be elaborated herein. Specifically, in the embodiment, the display unit of the terminal is a touch screen display, and the terminal further includes the memory and one or more than one program, wherein the one or more than one program is stored in the memory, and is configured to be executed by one or more than one processor. The one or more than one program includes instructions configured to execute the following operations:

initial model parameters of a CNN model to be trained are acquired, wherein the initial model parameters include initial convolution kernels and initial bias matrixes of convolution layers of respective levels, and an initial weight matrix and an initial bias vector of a fully connected layer;

a plurality of training images are acquired;

on the convolution layer of each level, convolution operation and maximal pooling operation are performed on each of the training images to obtain a first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix of the convolution layer of each level;

horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain a second feature image of each of the training images on the convolution layer of each level;

a feature vector of each of the training images are determined according to the second feature image of each of the training images on the convolution layer of each level;

each feature vector is processed to obtain a classification probability vector of each of the training images according to the initial weight matrixes and the initial bias vectors;

a classification errors is calculated according to the classification probability vector and initial classification of each of the training images;

the model parameters of the CNN model to be trained are regulated on the basis of the classification errors;

the process of regulating the model parameters is continued on the basis of the regulated model parameters and the plurality of training images until the number of iterations reaches a preset number; and

model parameters obtained when the number of iterations reaches the preset number are determined as the model parameters of the trained CNN model.

If the above is a first possible implementation, in a second possible implementation provided on the basis of the first possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the plurality of training images are acquired which includes:

a plurality of initial training images are acquired;

for each of the initial training images, a width-height ratio of the initial training image is kept, and the initial training image is processed to obtain a first image with a specified height; and

the first image is processed to obtain a second image with a specified width, and the image with the specified height and the specified width is determined as the training image corresponding to the initial training image.

In a third possible implementation provided on the basis of the first possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the plurality of training images are acquired which includes:

a plurality of initial training images are acquired; and

for each of the initial training images, a width-height ratio of the initial training image is kept, the initial training image is processed to obtain an image with a specified height, and a width corresponding to the specified height is determined as width of the initial training image.

In a fourth possible implementation provided on the basis of the second possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the first image is processed to obtain the second image with the specified width which includes:

when the width of the first image is smaller than the specified width, left and right sides of the first image are uniformly filled with pixels having a specified gray-scale value, and the second image is obtained when the width of the first image reaches the specified width; and

when the width of the first image is larger than the specified width, pixels on the left and right sides of the first image are uniformly cropped, and the second image is obtained when the width of the first image reaches the specified width.

In a fifth possible implementation provided on the basis of the first possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the convolution operation and the maximal pooling operation are performed on each of the training images to obtain the first feature image of each of the training images on the convolution layer of each level by using the initial convolution kernel and initial bias matrix on the convolution layer of each level which includes:

for each of the training images, the first feature image on the convolution layer of a previous level is input to a current convolution layer, and the convolution operation is performed on the first feature image on the convolution layer of the previous level to obtain a convolutional image on the current convolution layer by using the initial convolution kernel and initial bias matrix of the current convolution layer, wherein the first feature image on the convolution layer of the previous level is the training image if the current convolution layer is the convolution layer of the first level; and

after the maximal pooling operation is performed on the convolutional image on the current convolution layer to obtain the first feature image of the training image on the current convolution layer, the first feature image on the current convolution layer is continuously transmitted to the convolution layer of a next level, and the convolution operation and the maximal pooling operation are performed on the convolution layer of the next level until the convolution operation and the maximal pooling operation are performed on the convolution layer of a last level to obtain the first feature image on the convolution layer of the last level.

In a sixth possible implementation provided on the basis of the first possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain the second feature image of each of the training images on the convolution layer of each level which includes:

for the first feature image of each of the training images on the convolution layer of each level, a maximum value of elements of each of rows of each of the images in the first feature image on the convolution layer are extracted, wherein the first feature image includes a preset number of images, and the preset number is the same as each of the numbers of the convolution kernels and bias matrixes of the convolution layer;

the maximum values extracted from all the rows of each image are arranged into a one-dimensional vector according to arrangement of pixels of each image; and

the one-dimensional vectors of all the images in the first feature image on the convolution layer are combined to obtain the second feature image on the convolution layer.

In a seventh possible implementation provided on the basis of the sixth possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the feature vector of each of the training images are determined according to the second feature image of each of the training images on the convolution layer of each level which includes:

for each of the training images, elements of all rows of the second feature image of the training image on the convolution layer of each level are connected head to tail to obtain the feature vector of the training image.

In an eighth possible implementation provided on the basis of the first possible implementation, the memory of the terminal further includes instructions configured to execute the following operation that the classification error is calculated according to the classification probability vector and initial classification of each of the training images which includes:

the initial classification of each of the training images are acquired;

the classification error of each of the training images is calculated according to the classification probability vector and initial classification of each of the training images by using the following formula:

Loss=−Lny_label,

where Loss represents the classification error of each of the training images, label represents the initial classification of each of the training images, y_irepresents a element of the classification probability vector of each of the training images, and y_labelrepresents a classification probability corresponding to the initial classification; and

a mean of the classification errors of all the training images is calculated, and the mean of the classification errors is determined as a classification error.

In a ninth possible implementation provided on the basis of any one of the first to eighth possible implementations, the training images are images in a natural scene, the images in the natural scene include characters in different languages, and the CNN model to be trained is a language recognition classifier.

In a tenth possible implementation provided on the basis of any one of the first to eighth possible implementations, the CNN model to be trained includes four levels of convolution layers and two fully connected layers, and the convolution layers of the respective levels include the same or different numbers of convolution kernels and bias matrixes, the memory of the terminal further includes instructions configured to execute the following operation:

the operation that the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of at least one of the levels to obtain the second feature image of each of the training images on the convolution layer of each level including:

the horizontal pooling operation is performed on the first feature image of each of the training images on the convolution layer of a second level, the first feature image of each of the training images on the convolution layer of a third level and the first feature image of each of the training images on the convolution layer of a fourth level to obtain the second feature image of each of the training images on the convolution layer of the second level, the second feature image of each of the training images on the convolution layer of the third level and the second feature image of each of the training images on the convolution layer of the fourth level, respectively; and

the operation that the feature vector of each of the training images are determined according to the second feature image of each of the training images on the convolution layer of each level including:

for each of the training images, the feature vector of the training image are determined according to the second feature image of the training image on the convolution layer of the second level, the second feature image of the training image on the convolution layer of the third level and the second feature image of the training image on the convolution layer of the fourth level.

According to the terminal provided by the embodiment of the present disclosure, after the convolution operation and the maximal pooling operation are performed on the training images on the convolution layer of each level, the horizontal pooling operation is further performed on the feature images obtained by the maximal cooling operation. During the horizontal cooling operation, the feature images which identify features of the feature images in a horizontal direction may further be extracted from the feature images to ensure that the trained CNN model may break limits of a size of an image to be recognized in the horizontal direction during image classification identification and recognize an image to be recognized with any size, so that an image recognition application range of the CNN model trained in such a manner is relatively wider.

The embodiment of the present disclosure provides a computer-readable storage medium, which may be a computer-readable storage medium included in a memory in the abovementioned embodiment, and may also be a computer-readable storage medium existing independently and unassembled in a terminal. The computer-readable storage medium stores one or more than one program, and the one or more than one program is configured for one or more than one processor to execute a method for training a CNN model, the method including: