Research On Clothing Image Retrieval Based On Attention Mechanism And Feature Fusion

With the rapid development of the Internet,online shopping has become a common way of shopping.Users can choose their favorite clothes on the shopping platform.However,due to the diversity of clothing types and different classification standards,the text description of clothing by consumers and merchants is usually inconsistent,making it difficult for consumers to find the clothing information they need quickly and accurately,which greatly affects consumers’ shopping experience.Using image-based clothing retrieval technology can not only eliminate the cognitive bias of consumers and merchants on clothing,but also greatly reduce the time cost of merchants,and provide consumers with a more convenient and comfortable online shopping environment.In this context,this paper studies clothing image retrieval,and the main work is as follows:(1)Aiming at the problem that the network model is too complex and has high requirements for equipment,this paper proposes m F-Mobilenet network model integrating shallow,middle and deep features on the basis of lightweight network Mobile Net V2.To solve the problem that a single feature cannot fully express the information of clothing image,this paper proposes a clothing feature extraction method combining SPoC algorithm,Ge M algorithm and RMAC algorithm.Firstly,m F-Mobilenet network is used to extract features from clothing images.Secondly,SPoC algorithm,Ge M algorithm and RMAC algorithm are used to pool the feature.Then,the full connection layer is used to transform the dimension of the features and L2 norm is used to normalize the features.Finally,the normalized features are fused and used for garment image retrieval.Experimental verification was carried out In Deep Fashion’s In-shop and consumer-toshop data sets,and the accuracy reached 95.32% and 32.46% respectively at top-5.(2)To solve the problem that there is a big difference between the user’s selfie picture and the merchant’s clothing picture of the same style due to the influence of lighting,background and deformation factors,the SE-Rep VGG network model based on the attention mechanism is proposed,and the SPoC algorithm is used to process the features.Firstly,SERep VGG network model is used to extract clothing image features.Secondly,the SPoC algorithm is used to process the features and get the SPoC features based on attention perception.Then,the full connection layer and L2 norm are used to transform and normalize the features.Finally,the normalized feature is used for garment image retrieval.Verified by Deep Fashion’s In-shop data set and consumer-to-shop data set,the accuracy reached 95.94% and 42.36% respectively at top-5.(3)Based on the above two clothing image retrieval algorithms,the design and implementation of clothing image retrieval system software.The system software is developed using B/S three-tier architecture,My SQL database is used to save information,Html framework is used to write front-end web pages,Flask framework is used to write back-end applications,and finally user management module and clothing image retrieval module are realized.