In a possible implementation manner of the third aspect, the steps performed by the processor also include:

- obtaining an image sample set, and taking the image sample set as the training sample set;
- constructing a self-supervised learning pre-training network, performing the self-supervised contrastive learning on the self-supervised learning pre-training network according to the training sample set, to generate the learned feature extraction network;
- constructing the convolutional neural network model according to the learned feature extraction network;
- obtaining downstream training samples having labeled data; and
- performing a supervised learning on the convolutional neural network model by using the downstream training samples having labeled data, to obtain the trained convolutional neural network model.

In a possible implementation manner of the third aspect, the steps executed by the processor specifically include:

- constructing the self-supervised learning pre-training network, where the self-supervised learning pre-training network includes an online encoder, a momentum encoder and a loss calculation model;
- performing a transformation process on each image sample respectively, to obtain a first view V_i1and a second view V_i2corresponding to the image sample; inputting the first view V_i1corresponding to the image sample into the online encoder respectively, to obtain a first projection feature atlas corresponding to the image samples;
- inputting the second view V_i2corresponding to the image sample into the momentum encoder respectively, to obtain a second projection feature atlas corresponding to the image samples;
- obtaining a feature extraction network model by training according to the loss calculation model, the obtained first projection feature atlas corresponding to the image sample, and the obtained second projection feature atlas corresponding to the image sample.

In accordance with a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, the method for the quality detection of tunnel lining through ground-penetrating radar based on self-supervised learning as described in the first aspect is implemented.

In accordance with a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, when the computer program is executed by a processor, the method for the quality detection of tunnel lining through ground-penetrating radar based on self-supervised learning as described in the first aspect is implemented.

The present application relates to an intelligent automatic recognition technology for quality detection data of a ground-penetrating radar tunnel lining structure, in particular to a method for deep neural network training based on a large amount of unlabeled data by means of the artificial intelligence self-supervised contrastive learning technology, which can save the high time cost caused by manual labeling of a large number of ground-penetrating radar detection data, and can greatly improve the efficiency of the ground-penetrating radar artificial intelligence automatic recognition technology. Meanwhile, due to an increase in the number of training samples to be used, the recognition accuracy of the method provided by the present application is significantly improved compared with the conventional artificial intelligence recognition methods.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a schematic flow chart of a method for a quality detection of tunnel lining with a ground-penetrating radar based on self-supervised learning provided by the present application;

FIG. 2 is a schematic diagram of a pre-training network based on self-supervised learning provided by the present application;

FIG. 3 is a grayscale image of ground-penetrating radar measurement data;

FIG. 4 is a diagram of a recognition result of steel arches;

FIG. 5 is a diagram of a recognition result of voids;

FIG. 6 shows a curve of self-supervised training loss;

FIG. 7 shows curves of downstream task learning rate (lr) and loss function (loss); and

FIG. 8 shows a curve of average recognition accuracy.

DETAILED DESCRIPTION OF EMBODIMENTS

In order to make the problems to be solved, solutions and beneficial effects of the present application more clearly, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described here are intended only to explain the present application, and are not intended to limit the present application.

The deep learning technology based on convolutional neural network is a supervised learning technology, and the recognition accuracy of deep learning technology depends entirely on the number of learning samples and the quality of labels. In the process of network training, it is necessary to manually label the ground radar detection images to make a training sample set. Due to the high time cost of manual labeling, the efficiency and accuracy of the model training are low. In response to these problems, the present application applies a large amount of unlabeled data for deep neural network training by means of artificial intelligence self-supervised contrastive learning technology, which can not only save the time cost of manual labeling, but also improve the efficiency of ground-penetrating radar artificial intelligence automatic recognition technology.

The present application relates to an intelligent automatic recognition technology for quality detection data of a ground-penetrating radar tunnel lining structure, in particular to a method for deep neural network training based on a large amount of unlabeled data by means of the artificial intelligence self-supervised contrastive learning technology, which can save the high time cost caused by manual labeling of a large number of ground-penetrating radar detection data, and can greatly improve the efficiency of the ground-penetrating radar artificial intelligence automatic recognition technology. Meanwhile, due to an increase in the number of training samples to be used, the recognition accuracy of the method provided in the present application is significantly improved compared with the conventional artificial intelligence recognition methods.

To be specific: it is difficult to obtain a large amount of real tunnel lining internal data containing defect signals based on ground-penetrating radar detection and difficult to label, and the deep learning model trained by ground-penetrating radar data generated by simulation or GAN is difficult to deal with real tunnel lining internal GRP data. Thus, it is difficult for the current deep learning method to be applied in the tunnel lining internal detection project.

According to characteristics of ground-penetrating radar data, the present application provides a self-supervised learning algorithm based on contrast constraints, self-attention for dense contrastive learning (SA-DenseCL), which can be used for internal detection of tunnel lining.

The main idea of the present application is that:

- 1) Firstly, a large number of unlabeled real GPR data is applied to pre-train the self-supervised learning pre-training network to obtain a feature extraction network (backbone);
- 2) The extracted feature extraction network (backbone) is constructed into a convolutional neural network model, and finally a small amount of labeled GPR data is applied to fine-tune the convolutional neural network model, to obtain a trained convolutional neural network model.

Experimental results show that when a limited and equal number of annotated tunnel lining GRP images are used for fine-tuning, the pre-training method successfully improves the detection performance of the convolutional neural network model for real tunnel lining internal GRP data. Among them, an average precision (AP) of steel bars reaches 0.96703, increased by 1.171%, the AP of void defects reaches 0.81043, increased by 8.494%, and the AP of the secondary lining division (thickness) reaches 0.94665, increased by 3.438%. Results of the model test are verified by performing a tunnel opening at the tunnel site, and the results show that the method can be successfully applied in the project for an internal quality detection of actual tunnel lining.

Referring to FIG. 1, the present application provides a method for a quality detection of tunnel lining through ground-penetrating radar based on self-supervised learning, which includes the following steps.

In step 1, an image sample set is prepared and obtained.

Each image sample in the image sample set is obtained in the following manner:

In step 1.1, a tunnel lining structure is detected through ground-penetrating radar equipment in a 2-dimensional or 3-dimensional array, to obtain raw data of a ground-penetrating radar measurement.

In step 1.2, the raw data of the ground-penetrating radar measurement is processed to obtain pre-processed data of the ground-penetrating radar measurement.

In step 1.3, the preprocessed data of the ground-penetrating radar measurement is processed to obtain a GPR grayscale image, and the GPR grayscale image is taken as the image sample.

In step 2, the image sample set is taken as a training sample set; where, each image sample in the training sample set does not need to be marked with an image-sample label.

In step 3, a self-supervised learning task is constructed, and the self-supervised contrastive learning is performed by using the training sample set, to generate a learned feature extraction network.

The step 3 specifically includes the following steps:

In step 3.1, a self-supervised learning pre-training network is constructed; the self-supervised learning pre-training network includes an online encoder, a momentum encoder and a loss calculation model, as shown in FIG. 2.

A structure of the online encoder is exactly the same as that of the momentum encoder. The online encoder includes a first backbone network backbone (1), a first global feature projection model, a first pixel-level local feature projection model and a first line-level local feature projection model. The momentum encoder includes a second backbone network backbone (2), a second global feature projection model, a second pixel-level local feature projection model, and a second line-level local feature projection model.

In this step, the global feature projection model is to perform a projection on an overall feature of a feature map after a feature extraction via the backbone network. The size of the feature map output by the backbone is 7×7. After being down-sampled by an adaptive average pooling layer, the size of the feature map becomes 1×1, and then the feature map enters two linear layers for dimension reduction. ReLU( ) is used as an activation function between the two linear layers, and finally, a feature vector with a dimension of 128 is output.

The pixel-level local feature projection model: which is configured to project pixel-level local features of the feature map after a feature extraction via the backbone network. After two 1×1 conv layers, the feature map is then divided into 49 unit-sized 128-dimensional feature vectors to achieve pixel-level feature comparison learning.

The line-level local feature projection model: which is configured to project feature correlation information between local positions of the ground-penetrating radar image. When projecting, the line formed by a row of pixels is used as a projection unit for projection.

In step 3.2, n image samples in the training sample set are input to the self-supervised learning pre-training network, which are respectively represented as: an image sample P₁, an image sample P₂, . . . , an image sample P_n, where each image sample is an unlabeled grayscale image.

For any image sample P_i, i=1, 2, . . . , n, the following operations are performed:

- In step 3.2.1, a transform process is performed on the image sample P_ithrough two random data enhancement approaches to generate a view V_i1and a view V_i2;
- Among them, the data enhancement approach includes: a random cropping, a random flipping, a contrast adjustment, a saturation adjustment, a Gaussian blur and a solarization operation.

In step 3.2.2, the view V_i1is input into the online encoder, and the following operations are performed.

The first backbone network backbone (1) performs the feature extraction on the view V_i1to obtain a feature map F_i1, and the feature map F_i1is taken as a first feature map F_i1.

The feature map F_i1is respectively input into the first global feature projection model, the first pixel-level local feature projection model, and the first line-level local feature projection model. The first global feature projection model is configures to perform a global feature projection on the feature map F_i1, to obtain a global feature projection map Q_g(F_i1). The first pixel-level local feature projection model is configured to perform a pixel-level local feature projection on the feature map Fi1, to obtain a pixel-level feature projection map Q_d(F_i1). The first line-level local projection model is configured to perform a line-level local feature projection on the feature map Fi1 through, to obtain a line-level feature projection map Q_r(F_i1).

Specifically, the first feature map F_i1is input into the first global feature projection model to obtain a first global feature projection map Q_g(F_i1). The first feature map F_i1is input into the first pixel-level local feature projection model for the pixel-level local feature projection to obtain a first pixel-level feature projection map Q_d(F_i1). The first feature map F_i1is input into the first line-level local feature projection model for the line-level local feature projection to obtain a first line-level feature projection map Q_r(F_i1). The first global feature projection map Q_g(F_i1), the first pixel-level feature projection map Q_d(F_i1), and the first line-level feature projection map Q_r(F_i1) are taken as a first projection feature atlas corresponding to each image sample.

The view V_i2is input into the momentum encoder, and the following operations are performed.

The second backbone network backbone (2) performs the feature extraction on the view V_i2to obtain a feature map F_i2, and the feature map F_i2is taken as a second feature map F_i2.

The feature map F_i2is respectively input into the second global feature projection model, the second pixel-level local feature projection model, and the second line-level local feature projection model. The second global feature projection model is configured to perform a global feature projection on the feature map F_i2through tl, to obtain a global feature projection map K_g(F_i2). The second pixel-level local feature projection model is configured to perform a pixel-level local feature projection on the feature map F_i2, to obtain a pixel-level feature projection map K_d(F_i2). The second line-level local feature projection model is configured to perform a line-level local feature projection on the feature map F_i2, to obtain a line-level feature projection map K_r(F_i2).

Specifically, the second feature map F_i2is input to the second global feature projection model to obtain a second global feature projection map K_g(F_i2). The second feature map F_i2is input to the second pixel-level local feature projection model for the pixel-level local feature projection to obtain a second pixel-level feature projection map K_d(F_i2). The second feature map F_i2is input to the second line-level local feature projection model for the line-level local feature projection to obtain a second line-level feature projection map K_r(F_i2). The second global feature projection map K_g(F_i2), the second pixel-level feature projection map K_d(F_i2), and the second line-level feature projection map K_r(F_i2) are taken as a second projection feature map set corresponding to each image sample.

In this step: the line-level local feature projection is performed on the feature map F_i1through the first line-level local feature projection model to obtain the line-level feature projection map Q_r(F_i1), specifically, the line-level feature projection map Q_r(F_i1) is the first line-level feature projection map Q_r(F_i1).

Specifically:

- 1) Three linear layers are configured, namely, a first linear feature extraction layer, a second linear feature extraction layer and a third linear feature extraction layer.
- 2) The feature map F_i1is divided into three linear regions, namely, a first linear region, a second linear region and a third linear region.

The feature map F_i1is respectively input to the first linear feature extraction layer, the second linear feature extraction layer and the third linear feature extraction layer;

The first linear feature extraction layer is configured to perform a feature extraction on the first linear region to obtain a feature vector B1.

The second linear feature extraction layer is configured to perform a feature extraction on the second linear region to obtain a feature vector B2.

The third linear feature extraction layer is configured to perform a feature extraction on the third linear region to obtain a feature vector B3.

- 3) a similarity Sim (B1, B2) between the feature vector B1 and the feature vector B2 is calculated.
- 4) the similarity Sim (B1, B2) is multiplied by the feature vector B3 to obtain an output vector, that is the line-level feature projection map Q_r(F_i1).

In step 3.2.3, the following formula is applied in a loss calculation model to calculate a loss value:

L
=

(

1
-
α
-
β

)

⁢

L
g

+

α
⁢

L
d

+

β
⁢

L
r

In this formula: L represents an overall loss value; L_grepresents a global loss value obtained by comparing the global feature projection map Q_g(F_i1) with the global feature projection map K_g(F_i2); L_drepresents a pixel-level loss value obtained by comparing the pixel-level feature projection map Q_d(F_i1) with the pixel-level feature projection map K_d(F_i2); L_rrepresents a line-level loss value obtained by comparing the line-level feature projection map Q_r(F_i1) and the line-level feature projection map K_r(F_i2); α represents a loss weight coefficient of the pixel-level loss value; and β represents a loss weight coefficient of the line-level loss value.

Taking the global loss value L_gas an example, a calculation approach is that:

The global feature projection map Q_g(F_i1) is used as a query q, and q is derived from the feature representation of different views of a same original GPR image. That is: the global feature projection map K_g(F_i2) is referred to as a positive key k⁺, and the other feature representations from different raw GPR images are considered as k⁻.

Therefore, for each query q, there is a set of encoded keys {k0, k1, k2 . . . }, but only one corresponding positive key k⁺ forms a pair of positive samples with the query q, and the remaining k⁻ forms many pairs of negative samples with the query q. A convergence direction of network training is to keep q close to k⁺ and away from all other k⁻. Based on this principle, a calculation formula for obtaining the loss function value is constructed.