Research On Virtual Try-on System Based On Image Semantic Features And Deep Hashing Features Retrieval

The Internet and e-commerce technology have achieved rapid development in recent years,and online shopping is becoming a common choice in people’s daily lives.Users can choose their favorite clothing on shopping platforms according to their preferences.At present,although online clothing shopping platforms have achieved tremendous development,there are still the following shortcomings: 1)In the online shopping mode,consumers are unable to predict their fitting effects based on graphic and textual information,resulting in high return and exchange rates;2)Virtual fitting technology can to some extent alleviate the pressure of a large number of returns and exchanges on online shopping platforms.However,the existing virtual fitting experience mainly relies on users’ active selection,and most lack the recommendation virtual fitting function for products of interest,as well as personalized and high-precision recommendation and fitting experience.Therefore,it is necessary to develop an online virtual fitting system with precise retrieval capabilities.In recent years,image retrieval methods based on convolutional neural network model semantic features can obtain better image Semantic information and higher retrieval efficiency than traditional algorithms.However,they lack the distinction between clothing target areas and background information in complex scenes,resulting in limited retrieval accuracy.Although deep hashing based clothing retrieval can achieve efficient clothing matching,its ability to distinguish different types of similar clothing still needs to be further improved.To address the above issues,this article designs a highly accurate and low latency virtual fitting system based on clothing image retrieval based on deep learning technology.It can solve the problem of existing virtual fitting systems only targeting clothing images that users actively select for virtual fitting,reducing user query time to a greater extent through retrieval,and enriching the shopping experience.The paper mainly optimizes image retrieval technology from two aspects: convolutional neural networks and deep hashing methods,providing users with more accurate clothing retrieval recommendation functions.At the same time,considering the deployment of devices,a lightweight model was constructed.The main research content is as follows:(1)Aiming at the problem of difficult spatial capture of clothing image position information in existing image retrieval methods due to factors such as complex backgrounds and small target areas.This article introduces a coordinate attention module in the Rep VGG structure,which strengthens the network’s attention to the spatial position information and long-distance dependencies of the target clothing,accurately locates the area where the clothing object is located,and ignores redundant surrounding background information.At the same time,in order to enhance the network’s access to the context information of clothing images and improve the utilization of features,a two branch hole convolution feature fusion structure FFM is constructed,which can obtain multi-scale Receptive field feature information of images.The improved Rep VGG method achieves classification accuracy of 96.5% and retrieval accuracy of 85.71% on the In shop clothing retrieval dataset,respectively,which is 1.6% and 2.45% higher than the original algorithm.(2)The existing deep hash algorithms seldom use the global supervision information of the data set,have insufficient constraints on the compactness and separation of the hash codes generated from samples with difficulty in classification,and do not make full use of the Semantic information loss caused by the category level semantics of the generated hash codes.An asymmetric center similarity hash algorithm is proposed.The loss function of the algorithm combines contrast loss,center similarity loss and category loss based on label smoothing.Among them,the contrast loss can maintain the similarity relationship of the original image in the Hamming space,the category level loss can focus on the semantic space distribution in hash learning,enhance the potential relevance of similar hash codes,and the central similarity loss will guide the hash function to generate hash codes with more convergent distribution,effectively improving the discrimination ability of hash codes to difficult samples.The experiment shows that the proposed algorithm achieves a retrieval accuracy of m AP of 83.5% on the clothing retrieval dataset,which is 1.9%and 11.2% higher than the deep hash method ADSH and CSQ methods,respectively.(3)To address the issues of large parameter size and slow inference speed in the deployment model of deep hash algorithm on mobile terminals,a lightweight improvement was made using the Mobile Net-V2 backbone network.The improved model has a parameter size of 2.31 M,a decrease of 21.22 M compared to the Res Net50-ACSQH model,and a retrieval accuracy decrease of only 0.22%.To compensate for the accuracy loss caused by model lightweight,the ability to extract clothing semantic features was improved by introducing FPN fusion network and parameterless attention mechanism.The optimized Mobile Net-ACSQH model reduced the number of parameters by 78.6% compared to Res Net50-ACSQH,and improved accuracy by 1.53%,meeting the real-time and accuracy requirements of terminal devices.(4)Finally,a virtual fitting system based on clothing image retrieval was designed.The system functions include user management,clothing retrieval,clothing fitting,retrieval information statistics,etc.Implemented personalized recommendation function for virtual fitting system.