Online Clothing Retrieval And Recommendation Based On Depth Network

It has been very popular that online shopping and selection of goods in the Internet era.The online market has become the main battle of the major clothing electric business competition.Manufacturers will introduce some clothes show video for users to watch online,and video in the clothing is generally recommended based on the user's searching records.How to automatically recommend the user clothing and find the clothes in library which are needed by users accurately,is a still less research topic.This paper is devoted to the study of clothing positioning,similarity matching,recommendation algorithm and model acceleration in online clothing video.We locate the upper body,lower body and whole body clothing of models in video precisely.Then the feature of the located clothes is extracted.Finally we use the extracted feature to match the clothed in database precisely.At the same time,this paper also set up a set of scientific test standards which is used to compare the video matching algorithm performance.For clothing location and keypoints positioning,we locate the upper body,lower body,the whole body type clothes using faster RCNN.After detecting the clothes position,we extract the clothing features near the feature points which are located using clothes alignment,landmark detection and ROI pooling.Finally,we fuse the feature extracted by the global branch and the local branch.For feature extraction,the residual unit and the attention unit are involved in the depth learning network,so that the algorithm can shield the invalid image area and concentrate the feature reprasentation of the garment itself.At the same time,we use metric learning to make the algorithm more accurate,train a very compact feature space by Triplet Loss,so that the similarity degree of similar samples is higher.It can greatly improve the accuracy of feature matching.In order to speed up the model,this paper proposes a model acceleration method based on mimick,which accelerates the whole matching system,so that the algorithm can realize real-time calculation in online video.This paper builds an online clothing recommendation system which can be applied in real time video,and designs a network which can be end-to-end trained,which has high academic and commercial value.