Cloth Detection And Matching Method Based On Multi-task Deep Convolution Network

With the development of the clothing e-commerce and the coming of the era of big data,a growing number of offline clothing stores began to sale clothing in the electronic commerce platform.Huge amouts of clothing data stores in the Internet and users draw more attention to their own clothing collocation.With huge amounts of data and algorithm of recommending clothing collocation to satisfy the demands of users and improve the user’s purchase desire,is an important means to improve the user experience and sales for clothing electric business platform.In addition,in modern society,people regard what they wear as the expression of their personality,and they also pay more attention to their clothes match.So dress collocation technique is of vital significant for both electric business platform and the user.Based on the application background,deep learning network model and time series model is applied to achive clothing collocation in this paper.The main work is following:1.The residual structure of recursive fusion module is proposed to improve the semantic segmentation accuracy as well as reduce parameters.The recursive structure reduces the parameter through reuse the convolution kernel,and the residual structure relieves the difficulty of gradient vanishing,and the feature fusion structure improves the segmentation performance through merging feature maps of different sizes of field.The module is also applied to the the segmentation branch of Mask RCNN network,and improves its segmentation performance.2.The clothing collocation network based on Bi – LSTM is proposed,and there are three modules.A)feature extraction module,respectively,using the Inception V3 to extract low-level feature information and the classification brank of Mask RCNNto extract high-level semantic information;B)compatibility module regard clothing pieces as time series,and Bi-LSTM is applied to study the relationship between pieces and the integral compatibility;C)feature fusion module introduces reference vector which is the fusion of two kind of image feature as the expression of an image,and it can express the image information better.The design of loss function can not only improve the similarity of the two kinds of image features in one image,but also increase the diversity between the two image.The experimental results show that the proposed loss function and the reference vector has promoting effect in the task of filling in blanks and scoring clothing collocation.3.This paper propose an end-to-end clothing collocation system,an image is segmentated into clothing item and then is sended to clothing collocation network after post-process.The whole system is made up of early instance segmentation and later clothing collocatio.This system is an end-to-end system and an image which contains clothing is required to complete clothing matching recommendation and score this suit,and it can be easily applied to Internet platform for users to find the suitable clothing collocation solution.