Research On Clothes Image Recognition,Landmark And Retrieval Based On Deep Learning

In recent years,the research on clothing image has attracted more and more attention with the development of computer vision technology in the field of deep fashion.At present,in the field of visual fashion analysis,the research mainly focused on clothing image recognition,landmarks detection,clothing retrieval,recommendation and matching.These studies could provide support for the decision of the production,manufacturing,inventory,sales and other links and bring a new shopping experience.However,in the actual application scenarios of clothing image,there are still challenges brought by the changes of clothing style,texture,tailoring,deformation,environment,placement and posture,which make the effect of recognition,positioning,recommendation and other applications not ideal.Based on deep learning methods,this paper mainly studied clothing recognition,landmarks positioning and clothing retrieval,and completed the following work:(1)This paper summarized the domestic and foreign research results and current research situation of image recognition and deep learning,analyzed the research direction of clothing image,and mainly studied clothing image recognition algorithm and clothing key point detection algorithm based on deep learning.On the basis of classic network structure,this paper respectively constructed the clothing recognition network and clothing key point detection network for the single task of recognition and detection,and compared the performance of the algorithm based on the evaluation indexes of m AP and MSE.(2)On the basis of the above research,this paper aimed at the influence of occlusion,deformation,illumination and complex pose of clothing image on recognition accuracy and key point positioning accuracy.Combining with SE module,void convolution and feature fusion thought,this paper combined the two single tasks,which were the clothing recognition and key point detection together,and constructed a multiple task network which could perform clothing recognition and key point detection simultaneously.This network adopted hole convolution instead of down sampling operation to increase the receptive field of the feature image.At the same time,the network used feature fusion method to fuse the image features extracted at different stages to increase the multi-scale expression ability.Through the experiment and analysis on the verification set,the results showed that compared with the single task network,the recognition accuracy of Top-3 and Top-5 clothing on the evaluation index of m AP improved by 3.1% and 2.5% respectively,and the accuracy of MSE key point evaluation index decreased by 2.2%.The experiment showed that the multi task network combined with feature fusion could effectively enhance the ability of feature expression and improve the performance of clothing recognition and positioning.(3)This paper used the above trained convolution neural network model to extract clothing features for image retrieval.In order to eliminate the influence of complex background in the experiment,this model used Faster-Rcnn algorithm to detect the main body of clothing.Taking the retrieval database image as the feature database,the model conducted similarity calculation and sorting between the image feature to be retrieved and the feature database to complete the image retrieval task.And on the basis of this method,this paper developed a retrieval system for clothing image,which could be used for completing retrieval tasks of similar clothing image.