Research On Clothing Multi-Attribute Recognition Method Based On Multi-Model Fusion Network

Today,with the rapid development of information technology,the quantity of garment goods is more and more large,and the types and styles are more and more.The rapid development of digital garment industry has brought the positive side to consumers,but also to a certain extent for consumers: too many choices,choice itself has become a burden.Increasing the efficiency of garment search and adding personalized recommendation has become an important part of improving consumers’ online garment shopping experience.However,clothing,as a special commodity,has the characteristics of softness,flexibility,variety and so on,which leads to the problems that the recognition effect is affected and the result is single.Aiming at the above problems,this paper proposes a method of clothing recognition based on multi-model fusion network,and studies the multi-attribute recognition of clothing.This article carries on the following research work:(1)A study of garment attributes is conducted to obtain multiple characteristics of garments by analysing the different semantic information they contain in their visual communication.The characteristics possessed by colour,style and category of garments are mainly studied in depth,and an objective method for the hierarchical labelling of garment datasets is proposed.(2)To solve the problems in the garment recognition task,a garment recognition method based on multi-model fusion network is proposed,which extracts global and local features of garments through Efficient Net and Res Net-18 models respectively,fuses the multi-features,aims to improve the recognition accuracy under complex influence conditions.The Efficient Net and Res Net-18 network models are investigated and the model structure is improved to address known shortcomings.The Fused-MBConv module is proposed to replace the MBConv module in the shallow structure of the original model in order to reduce the number of model parameters and increase the training speed.The introduction of the SK module for adaptive model tuning is aimed at increasing the ability to extract features.In the Res Net-18 network model section,the model structure is lightened and improved to reduce the model complexity of the multi-model fusion network.(3)To solve the problem of single recognition result,this paper researches on multi-label learning and multi-task learning,and selects the method suitable for multi-attribute recognition task in this paper through experimental comparison.And based on the above results the clothing recommendation application implementation method is explored and implemented,using clothing elements and collaborative filtering methods based on clothing recommendations,and building a WEB system aimed at verifying the feasibility and practicality of the clothing recommendation part.