Clothing Style Identification And Retrieval Based On Deep Learning

Clothing style is an important attribute in clothing image,which describes the overall sensory effect of a clothing image.Different regions have different definitions and divisions of clothing styles.For example,The Japanese style divides the clothing into lolita,classical,fairy and other categories,while the European style divides the clothing into Gothic,Bohemian,fashion and other categories.Since non-professionals do not have relevant knowledge in the field of clothing,it is difficult to conduct style identification and retrieval of clothing images.With the development of computer vision technology,the pattern of image searching has become an important way to search clothing images.The study on how to make the computer automatically recognize and retrieve the clothing style can help the ordinary users who do not have the professional knowledge of clothing to quickly search the image of goods containing their favorite clothing style.Based on deep learning technology,this paper studies the identification and retrieval of clothing style,designs the improved Bilinear-CNN clothing style identification model and Bilinear-DSH clothing style retrieval model respectively,and designs and implements a clothing style identification and retrieval system for ordinary users.The specific work is as follows:1.Style identification of clothing images.To solve the problem that the original Bilinear-CNN model is easily affected by the background of the garment image,the spatial attention mechanism is introduced in the deep convolutional layer to make the model pay more attention to the part of the garment in the image.Aiming at the problem that the original Bilinear-CNN model has a large number of parameters and computation,a single channel bilinear pooling method is designed,and the global maximum pooling and global average pooling are adopted to calculate the Gram matrix,so that the model can not only extract the feature correlation information,but also reduce the number of parameters and computation.The experimental results show that the improved Bilinear-CNN model has high recognition accuracy and speed.2.Style retrieval of garment images.To solve the problem of time-consuming similarity calculation between high-dimensional features,a Bilinear-DSH hash retrieval model was designed.The feature extraction part of the Bilinear-CNN model and the hash coding part of the DSH model were fused,and the extracted fine-grained features were mapped to a binarized hash space.Aiming at the problem that the super parameters in Bilinear-DSH model affect the retrieval performance,bayesian optimization algorithm is used to optimize the super parameters.Experiments show that the Bilinear-DSH designed in this paper can achieve high retrieval accuracy in a short time.3.Realization of clothing style identification and retrieval application.A resize-padding pretreatment method is designed to solve the problem of user input image size diversity.The image size is standardized while keeping the garment size constant.In order to better display the embodiment of clothing style attributes on user images,class activation mapping algorithm is used to visualize the position of style features.This system is aimed at ordinary users who do not have professional knowledge of clothing.It is hoped that they can identify themselves or their favorite style of clothing through this system and carry out retrieval.The whole system only needs the user to input an image and cut it,then the garment style recognition and retrieval task can be completed.Therefore,the system has a high degree of automation and has a great application prospect.