Research On Fast Clothing Image Retrieval

With the popularity of internet,the online shopping has become an important approach of daily shopping.However,the online clothing has many categories and styles.It is difficult to describe clothing with text tags.This brings trouble of online shopping.The online clothing image retrieval comes into being.Through searching by image,we can find the similar clothing image accurately,which can improve the efficiency of online shopping.There are many difficulties on online clothing retrieval in the following reasons: 1.numerous clothing image attributes,such as a variety of materials,complex texture,different colors etc.2.Complex background in daily clothing image.Clothing is susceptible to occlusions,illumination and so on 3.Clothing is non-rigid object which is prone to deformation.4.There are billions of clothing images on line,it is difficult to retrieve images efficiently.Currently,online clothing retrieval has got a lot of attention of scholars on machine learning and other related fields and has made great progress.But there are still many problems.In order to solve the existing problems in clothing retrieval,this paper presents a new segmentation and retrieval method,the main research content is as follows: 1.This paper investigates advanced clothing retrieval methods in the world and describes the related technologies in detail.We analyze the advantages and disadvantages of current clothing segmentation and retrieval methods.Aiming at the problems,we proposed new approaches and improved the process of clothing retrieval;2.Aiming at present problems of clothing segmentation method,this paper proposed semantic segmentation using fully convolutional network to remove the complex background and maintain the human region.We combine the results of pose estimation and super-pixel segmentation to segment the complete clothing accurately;3.In the view of clothing representation,clothing products own multiple attributes.In order to represent the characteristics of clothing,this paper proposed computing different features of clothing and then using multi-model learning to learn the relationship of different features to represent the characteristic of clothing and obtain efficient clothing representation.This method can improve the ability of representation,recline the dimension of features and improve storage and retrieval efficiency;4.This paper studies fast retrieval on deep feature of clothing.Because the high dimension of deep feature leads to low retrieval effectiveness,we introduce bag-of-words model into deep feature,transforming deep feature into word frequency histogram,making deep feature sparse and combining it with inverted index to improve the retrieval efficiency.Finally,we compare our method with the current state-of-the-art clothing retrieval approaches.The experiment verified that our method have better results on clothing retrieval.