Clothing Image Classification And Retrieval Based On Deep Neural Network

With the development of e-commerce,online shopping has become a fashionable way of shopping.Clothing images classification and retrieval have great research value because of their unique application prospects.Faced with the increasing number of clothing images,how to quickly and accurately classify and retrieve clothing images is a meaningful research.At the same time,due to the interference of complex background,the research of clothing images is facing severe challenges.How to reduce the interference of the background appropriately and improve the accuracy of the classification and retrieval is a recent research trend recently.Therefore,this thesis proposes a method of image retrieval based on clothing area.In this thesis,a new method of attribute classification and image retrieval for clothing images is proposed.The main work is as follows:Firstly,the basic theoretical knowledge of deep learning is elaborated,its working principle is analyzed,and the related algorithms and optimization problems are studied.Aiming at the strong position correspondence between some attributes of clothing image and the middle part of the image,such as the attribute of sleeve length is determined by the area of sleeve,a method of attributes classification based on clothing image is proposed.Firstly,the method combines the principle of pose estimation to determine the location of some regions related to attributes in the image.Secondly,the deep neural network model is used to extract the local features of the relevant regions and the global features of the whole image.Finally,the global features and local features are fused as the final classification features of the attributes of the clothing image.At the same time,the fusion features obtained by this method are compared with the single global features of the image,and experiments are designed to explore the influence of different depth of the neural network model on the method.Experiments show that it is effective to apply pose estimation to attribute classification,and can quickly and accurately obtain the relevant regions.Moreover,the fusion features obtained by this method perform well in any kind of neural network model.However,compared with the shallow neural network,the deep network will perform better,which shows that the features acquired by the deep neural network in this method are also beneficial to image feature representation.When the classification is accurate,users can search according to classified good items.In the task of image retrieval,this thesis proposes a method of image retrieval based on clothing area.This method is based on pose estimation.According to the specific area obtained by pose estimation,clothes are located in the image.At the same time,the algorithm of Ro I Align in target detection algorithm is introduced to obtain the mapping clothes area feature map.Finally,the mapping feature map and the whole feature map are fused.The fused feature is used as the retrieval feature.Online triplet loss is trained as a loss function.This method integrates the Ro I Align concepts of target detection and pose estimation into image retrieval,and is dedicated to the determination of local regions,the extraction of local features and the mapping of features.Finally,this method is compared with WTB,DARN and Ro I Pooling methods to compare the retrieval performance under different networks.The results of experiment show that the method has high accuracy,which can reduce the interference of background on image retrieval,and can improve the image retrieval performance.