Clothing Retrieval Based On Deep Learning

As the development of mobile Internet and E-commerce,image retrieval,especially clothing retrieval technology,has drawn increasing attention from both industry and academy in recent years.This paper focuses on several technology of clothing retrieval.Based on the deep learning model,we investigate some key issues for the clothing retrieval.This paper mainly deals with fallowing problems for clothing retrieval: 1.Image preprocessing.In order to get rid of the influences of background,gesture,etc,we need to identify the clothing region first.2.Supervised classification and feature extraction.Different from traditional hand-craft feature extraction methods,deep learning model needs to train a deep neural network,and then extracts visual features from certain layer as the retrieval feature.3.Retrieval acceleration.As the increasing number of images in the search library and the dimensions of feature vectors,the time cost of retrieval will increase dramatically,which calls for an efficient clothing retrieval algorithm with satisfactory precision in the meanwhile.The main contributions and innovations are as follows: 1.To adapt the clothing search problem,we modify the general framework of Faster R-CNN for clothing region detection.In particular,we propose a priori based stochastic sliding window algorithm to produce the proposal windows instead of using the region proposal network(RPN),which simplifies the training phase and improves the detection precision.2.Considering the multi-attributes characteristic of clothing,we propose the hierarchical tree structure deep neural network which can improve the precision of of high-level semantic attributes.In addition,the proposed network can learn the fused feature of low-level visual features,middle-level semantic features and high-level semantic features to refine the retrieval feature.3.We use coarse filter and fine-grained retrieval,and the principle component analysis(PCA)algorithm to accelerate the retrieval phase.We crawl a lot of clothing images from several E-commerce websites like JD,Taobao,etc.Besides images,we also crawl the text description of clothing attribute provided by shop keepers,and then make them structured in the database.With the dataset at hand,we conduct experiments on clothing region detection,clothing classification,clothing retrieval,and retrieval acceleration.The results of experiments show that our clothing detection module is accurate and efficient.The results also show that our hierarchical tree structure network can improve the accuracy of high-level semantic attributes and can learn representative features of clothing image.Furthermore,it turns out that our acceleration method reduces the time cost while keeping the good performance of retrieval.