Research On Clothing Image Retrieval Based On Deep Learning

With the booming development of clothing e-commerce in recent years,online clothing image data has exploded.How to quickly and accurately retrieve clothing images of interest to users from a large collection of clothing images is a challenging and promising application.Traditional clothing image retrieval based on a large amount of label information is time-consuming and labor-intensive,subjective,and it is difficult for label information to fully and accurately describe certain characteristics of clothing,making the retrieval results difficult to satisfy.The online clothing search mode of "Pisoutu" is essentially a retrieval technology based on the characteristics of image visual content.There are still problems such as feature selection and semantic gap.Since deep learning has achieved great success in computer vision and other fields in recent years,this article introduces deep learning technology into the field of clothing image retrieval,and researches clothing image retrieval based on deep neural networks from two different perspectives.First of all,the shallow and deep features of the deep neural network model are used for clothing image retrieval.A clothing image retrieval method combining multi-level feature fusion of convolutional neural network and K-Means clustering is proposed.In the deep convolutional network structure,the detailed information of the clothing object is obtained by using the shallow network structure,and the overall information of the clothing object is obtained by using the deep network structure,so as to obtain a clothing image feature vector with more expressiveness,and realize Improved retrieval accuracy.Specifically using GoogleNet as the benchmark network,network features are extracted from three different Inception module groups,and multi-level features of clothing images from low to high are extracted,and then multiple levels of features are merged as the final clothing feature representation vector.At the same time,in the similarity measurement stage,in order to overcome the problem of time-consuming distance measurement of clothing picture vectors one by one,K-Means clustering is used to first divide the clothing image into K clusters,and then the vector of the image to be retrieved and each cluster The vector in the center performs similarity calculation,determines the category of the picture to be retrieved,and narrows down the scope of retrieval.The proposed method is verified on some data in the DeepFashion dataset,and compared with traditional HOG and HSV feature extraction methods and direct extraction of GoogleNet deep network features for experimental comparison.The proposed experimental scheme can effectively enhance the feature expression effect of clothing pictures and improve " "Search by map" is accurate,and the clustering method also greatly saves retrieval time and improves experimental efficiency.Secondly,the color image and deep network features of clothing images are combined to realize clothing image retrieval.A clothing image retrieval algorithm that combines color features and depth characteristics is proposed.Because color plays a very important role in clothing,but the problem of similar clothing styles but large color differences often occurred in clothing retrieval directly using deep learning methods.Therefore,deep neural networks were used to obtain color features and deep network features and fuse them for clothing images.Retrieve.The Resnet50 network in the convolutional neural network was used as the main network,and the corresponding improvements were made.The last convolution layer of the residual network was extracted to obtain the color features of the clothing image,and the deep network features of the clothing image were fused as the final features.The vector makes full use of the hierarchy and effectiveness of the residual network in image feature extraction.Validation was performed on some datasets of Deepfashion,FashionAI,and other e-commerce websites.The results show that this method can improve the color expression ability while extracting the global features of the deep network,and combine the various features of clothing pictures.The information improves the accuracy of clothing image retrieval on a certain basis,and has strong practicality.