Research On Garment Image Retrieval Based On Landmark Feature

With the development and prevalence of the internet and smart phone,online shopping has become one important part of daily life for today's people.All e-commerce platform do their best to provide a more convenient and comfortable online shopping environment for users.Coventionally,they search their target goods on the online shopping platforms by keywords based on text information.However,there is a huge amount of goods and its related information.It is usually difficult to obtain accurate results only by key words searching especially for those that has strong and rich appearance like clothing.Generally,the text description is more subjective,and cannot match the target clothing more precisely.Alternatively,the image-based clothing search methods are proposed to address the issue,and have attracted more attention from both the academia and industry.After receiving the characteristics of the clothing image input by the users,the online shopping platforms can match the clothing image and return result with the similar clothing and related online store.Additionally,the image-based clothing search strategy can make a recommendation of clothing products in a simple and effective way according to user demand.This thesis focuses mainly on the clothing image retrieval.We utilize the convolutional neural network to express the feature of clothing images with aiming to improve the performance of image retrieval.In this work,the local features are extracted based on the key points of the clothing to describe the information about the functional local areas of the clothing.The global features of the clothing image is extracted by convolutional neural network to describe the global information of the clothing image,which has certain advanced semantic information.We integrate local features of key points and global features of the clothing to jointly distinguish the clothing images,which can effectively improve clothing image retrieval performance.Since the local features of the clothing are extracted based on the location of the clothing key points,the accuracy of clothing key point positioning will directly affect the expression results of clothing functional local features.A clothing key point detection method based on pose estimation,which combines with the scheme of sequential reasoning machine and convolutional neural network,reveal the invisible spatial relation of different clothing key points.In this thesis,we utilize the deep neural network to extract the global features of clothing images and the local features of key points.In order to take advantage of different levels of information extracted from different depth convolutional layers,this thesis proposes to extract the local features of clothing key points based on multi convolutional layers,and it is able to make the features of clothing images become more fine-grained and discriminative.At the same time,we introduces Triplet Loss to learn the similarity measure for ensuring the rationality,which improve the accuracy of the clothing image retrieval task.This thesis firstly verifies the effectiveness of the clothing key point position method through experimental investigation,and the experiment results prove the accuracy of locating the functional area of clothing and expressing the local features.Meanwhile,this method can eliminate the impacts of the complex background of the clothing and cope with the deformation of the clothing.Based on the results of the experiments,combining with the local features and global features,the proposed strategy can effectively express the clothing features,and improve the performance of clothing image retrieval.As a result,the clothing image retrieval strategy provides a better recommendation of similar clothing,especially for users who pay more attention to clothes style.