Research On Related Technology Of Cross-scenario Clothing Image Retrieval

With the development of Internet software and hardware technology,Various E-commerce platforms flourish,the online shopping has become an important approach of daily shopping.Among them,clothing shopping has become one of the most popular applications.Therefore,it's an urgent problem to be solved that retrieving out quickly what customer need from vast number of clothing products.In response to this problem,cross-scenario clothing retrieval came into being.According to the input of clothing images,the required products should be researched quickly and accurately,the online shopping experience should be improved.Currently,cross-scenario clothing retrieval has got a lot of scholars' attention,however,there are still many problems,including the following: 1)The background of daily clothing image is complex,and the interference factors such as attitude and light introduce unnecessary noise for the follow-up work;2)There are too much clothing categories and styles,textures,colors and so on,have brought difficulties to the retrieval of clothing products;3)Clothing images belong to the low-level feature,requirement belong to the high-level semantic information.There exists big semantic gap between high-level semantic information and low-level clothing image features.According to the main problems encountered in technologies of the current cross-scenario clothing retrieval,we propose a new cross-scenario clothing retrieval method.The main research contents can be concluded as follows:1.The existing mainstream clothing retrieval algorithms at home and abroad comprehensively are analyzed,the related technologies are introduced and the advantages and disadvantages of these methods are find.Aiming at the defects of existing methods,we attempt to give some solutions and further improve the related fields of clothing image retrieval.2.On the basis of content-based retrieval,we propose a new cross-scenario clothing retrieval network framework.Adopted Faster R-CNN algorithm,we detect the clothing region.Then,we extract multi-level features and constraint retrieval process by the clothing class information.By this way,we improve the retrieval accuracy.3.In multi-level feature extraction,we first puts forward the imagine that the useful information of multi-scenario data not only contain the common features but also the specific information in different scene domains.Based on this imagine,A multi-level feature extraction model is constructed.This method can effectively improve the discriminating power of garment feature descriptors.4.An enhanced contrastive loss function based on class constraints on the basis of multi-level features is proposed.We constrained the traditional contrast loss function by additional class information to prevent over-fitting.Finally,the experiment results show that our method outperforms the state-of-the-art approaches.