Research On Few-sample Image Classification Based On Deep Metric Learning

Modern people’s life has become more and more inseparable from retail products,but a wide variety of products also bring a lot of trouble to people when shopping.The"new retail" that combines the traditional retail industry with e-commerce is becoming the future development trend of the retail industry.How to use artificial intelligence and big data to upgrade sales and distribution methods and improve online and offline shopping experience is an urgent problem to be solved in the new retailing.Taking offline shopping in retail stores as the application background,this thesis implements a few-sample commodity image classification system based on deep metric learning.This system can classify commodities by learning the similarity measure between commodities in the case of few samples,which can effectively reduce the training time and cost of the network,provide users with a new way to query product information,and improve the shopping experience in offline retail stores.The main research results of this thesis are as follows:Firstly,a product classification network based on deep metric learning is built on the basis of the Siamese neural network structure,and according to the characteristics of the retail product classification task,the corresponding feature extractor and product classifier are designed and constructed.In the feature extractor,a variety of different convolutional neural networks were tried for feature extraction,and a comparison experiment was conducted to compare the accuracy of commodity image classification;Secondly,this thesis designs a spatial structure optimizer for the problem that retail commodities contain many similarlooking commodities but belong to different categories.Through comparative experiments,it is proved that the optimizer can improve the accuracy of the network in commodity image classification;Finally,this thesis builds a few-sample commodity image classification system based on metric learning based on the network with the best comprehensive performance in the above experiments.