Clothing Retrieval And Outfit Composition Recommendation Algorithm Research Based On Deep Learning

The market scale of the apparel industry has shown a steady growth trend.Whether it is clothing e-commerce or the real economy,it has been developing at a high speed in recent years.Clothing commodity information in a state of explosive growth with the development of e-commerce,how to filter out useful information from the vast amount of clothing data,how to use this information to create higher value,has become an urgent problem to be solved in today's clothing e-commerce;At the same time,due to the impact of e-commerce,traditional real economy operation need to be reformed.How to make offline shopping more convenient and enhance the shopping experience of consumers is an urgent problem to be solved in promoting the development of the real economy.In order to solve the above problems,based on the deep learning theory,this thesis studies and implements clothing classification,retrieval and outfit composition recommendation methods.:First,we propose a multi-label classification method for clothing based on multi-task learning.Based on the SE module(Squeeze and Excitation),dilated convolution and diverse features fusion method,this thesis designs the SDD_RN network(Residual Network with Squeeze and Excitation,Dilated Convolution,Diverse features Fusion).In addition,based on the SDD_RN network,combines with clothing classification tasks for category,scale and design,this thesis proposes a multi-task learning method that utilizes data and task correlation to improve classification performance.Next,we propose a clothing retrieval method based on multi-feature fusion.Based on the categories,size and design tasks' deep features extracted by the multi-task classification network,and the color features,this thesis implements clothing retrieval based on multi-feature fusion.Compared with the method of single deep local feature retrieval,the method proposed in this thesis has higher retrieval accuracy.Combining visual semantic embedding and clothing style vector,this thesis proposes orderly and disordered personalized outfit composition recommendation methods.Based on the Bi-LSTM(Bilateral Long Short-Term Memory)network,this thesis constructs an orderly outfit composition network.The network regards the clothing combination as a top-down clothing sequence and realizes the outfit composition recommendation is implemented through sequence prediction.In addition,based on the Aggregated network,this thesis constructs a disordered outfit composition network.The network regards the clothing combination as a clothing collection and realizes the outfit composition recommendation by analyzing the distribution of clothing information.By extracting the embedded vector group of the user's interest library and using K-Means to cluster the user's style vector.Combined with the style vector for collocation prediction,this thesis can realize the personalized outfit composition recommendation of the clothing.Finally,based on the classification,retrieval and matching recommendation methods for clothing proposed in this thesis,we designed the intelligent clothing shopping guide system of physical store,which can be widely used in the clothing industry for smart shopping guide,information retrieval and outfit composition recommendation.