In one embodiment of the present disclosure, the intermediate embedding feature vector generation unit 208 may generate intermediate embedding feature vectors of different sizes based on the complexity of the respective fashion attributes. For example, because the task of distinguishing a category of an item is more complex than the task of distinguishing a color of an item, the intermediate embedding feature vector generation unit 208 may construct a larger intermediate embedding feature vector for the category than for the color.

According to one embodiment of the present disclosure, the single vector generation unit 210 may perform a function to generate a single vector that comprehensively represents an attribute for the corresponding image by combining the at least one intermediate embedding feature vector. Here, the single vector may be a vector of information on a fashion item included in the image, positioned in a style space, representing each attribute of the item. In one embodiment, the single vector may be generated by fusing each of the intermediate embedding feature vectors generated by the intermediate embedding feature vector generation unit 208.

According to one embodiment of the present disclosure, the single vector generation unit 210 may weight each of the intermediate embedding feature vectors based on the associated multi-task attribute. In one embodiment, the single vector generation unit 210 may weight each of the intermediate embedding feature vectors based on the complexity of the associated multi-task attribute. In one embodiment, the complexity of the associated multi-task attributes may vary depending on the attribute information corresponding to each component of the intermediate embedding feature vector, and the intermediate embedding feature vector generation unit 208 may weight the intermediate embedding feature vector based on such complexity. For example, the single vector generation unit 210 may weight the intermediate embedding feature vectors for item and category, which have more complex task attributes, more heavily than the intermediate embedding feature vectors for color, and fuse the intermediate embedding feature vectors to generate a single vector.

According to one embodiment of the present disclosure, the similar product determination unit 212 may perform a function of determining a visual search database image including a fashion item similar to a fashion item included in the query image as a similar product image. According to one embodiment of the present disclosure, the similar product determination unit 212 may compare a single vector of the query image with each single vector stored in the visual search database, and determine an image having the single vector that satisfies a predetermined criterion as a similar product image to the query image.

In one embodiment of the present disclosure, the similar product determination unit 212 may calculate a similarity between the query image and the visual search database images. According to one embodiment of the present disclosure, the similar product determination unit 212 may calculate the similarity by comparing a single vector of the query image to a single vector for each image in the visual search database. In one embodiment, the similar product determination unit 212 may calculate the similarity by comparing the distance of the single vector for the query image to the single vector for each image in the visual search database in style space.

In one embodiment of the present disclosure, the similar product determination unit 212 may select an original image for a single vector with a similarity that satisfies a predetermined criterion as a similar product image for the query image. In one embodiment of the present disclosure, the similar product determination unit 212 may select a single vector as the single vector for the similar product image if the similarity between the single vector in the visual search database and the single vector in the query image meets the predetermined criteria.

In one embodiment of the present disclosure, the similar product information provision unit 214 may perform a function of providing information of an item among images in the visual search database that includes an item similar to a fashion item included in the query image. In one embodiment of the present disclosure, the similar product information provision unit 214 may provide information of a fashion item included in the similar product image for the query image. In one embodiment, the similar product information provision unit 214 may provide the original image of the selected similar product image together with the sales information of the fashion item included in the image.

In one embodiment of the present disclosure, the similar product information provision unit 214 may sort the single vectors in the visual search database in order of increasing similarity with the single vector of the query image, and provide information on similar product images and the fashion items included in those images. In one embodiment, the similar product information provision unit 214 may provide information about similar products by sorting the similar product images for the query image by the popularity of sales on the online market. Alternatively, the similar product information provision unit 214 may provide information on similar products by various sorting criteria.

The communication unit 216 of the similar product search server 130 according to one embodiment of the present disclosure may perform a function to enable data transmission and reception from/to the search database management unit 202, the image receiving unit 204, the item location detection unit 206, the intermediate embedding feature vector generation unit 208, the single vector generation unit 210, the similar product determination unit 212, and the similar product information provision unit 214.

Similar product search server 130 according to one embodiment of the present disclosure may generate intermediate embedding feature vectors for attributes of fashion items included in original images and query images, by using a multi-task learning model sharing a unified backbone network, and fuse the intermediate embedding feature vectors into a single vector. According to the present disclosure, which uses a multi-task learning model to generate a single vector including various attribute features and uses it to search for similar products, it is advantageous to be able to use correlations between different fashion attributes and to check detailed features of a fashion item rather than merely overall features, compared to a method in the related art using a single task. For example, when comparing the similarity of items, the category, shape, and color features of the items may be compared collectively to determine similarity. A single vector including all of the different attributes, such as in one embodiment of the present disclosure, may correspond to any category of items, and more detailed attributes may be learned to improve similar product search performance. Further, according to one embodiment of the present disclosure, the similar product search server 130 has the advantage of easily scaling up as the search database grows by minimizing and weighting the intermediate embedding feature vectors for the attributes of the fashion items in accordance with the complexity of the multi-task attributes, that is, the task difficulty.

FIG. 3 is a flow diagram illustrating a process of configuring a visual search database according to one embodiment of the present disclosure.