Research On Clothing Matching Related Technology Based On Convolutional Neural Network

With the continuous development of artificial intelligence and online shopping,more and more people are bound up in online shopping.Among them,clothing is the largest part of online shopping.However,facing with the large amount of clothing image data with various of kinds and styles,how to match clothing items quickly and accurately is an urgent problem to be solved.To solve this problem,clothing matching came into being.Clothing matching is to match the clothing automatically which can meets the user’s need according to the clothing selected by users.However,most of the existing matching systems rely on the user’s interactive input.They extract the color,texture,and material features of the given clothing and match them mainly according to the similarity of their texture and material.On the one hand,the maching results can not satisfy people’s demand of clothing fashion.On the other hand,the feedback time is too long to meet the speed requirement for online matching.Aiming at the weakness of existing matching systems,this paper will study how to achieve clothing matching intelligently and quickly.The main contents of this paper are as follows:1.This paper expands a clothing dataset FClothes that meets fashion requirements.The FClothes clothing dataset was first built in 2016.Due to the changes of clothing fashion,this paper expanded the dataset in 2017.Based on the original clothing dataset,1000 outdated images were removed and 5600 new images were added.2.The detection and segmentation of clothing regions is an important pre-processing process for clothing matching.In this paper,Faster R-CNN technology is used to detect the clothing regions and an extra segmentation procedure is conducted for further segmentation according to the detection results,which can eliminate the influence of the complex background on clothing matching.3.In order to solve the huge time costing problem of clothing matching,this paper puts forward a novel learning fast style two-way network which combining Siamese CNN with hash coding.The hash feature and deep feature of the clothing images are first extracted through the proposed network,the hash feature is used for pre-matching,and the deep feature is used for re-matching.4.How to eliminating the cross-domain issue is the key to clothing matching.This paper uses Query Extension(QE)technology to eliminate the cross-domain problem of clothing matching.Finally,the results of comparison experiments with other clothing matching methods show the superiority of the proposed method.