Research On Clothing Keypoint Detection And Category Attribute Classification Based On Deep Neural Network

In recent years,with the rapid development of artificial intelligence technology and clothing e-commerce,new applications such as clothing image retrieval,personalized clothing recommendation and virtual try-on have received widespread attention,but these applications rely heavily on high-quality clothing keypoint detection and attribute classification.As a typical flexible object,the visual analysis of clothing is faced with many problems,such as complex background,variety of clothing,model posture and occlusion.These problems make the accuracy of keypoints detection and attribute classification of clothing not high enough to meet the needs of practical application.This paper aims to use the latest artificial intelligence technology,namely deep neural network method,to study the keypoint detection and attribute classification of clothing.(1)Aiming at the occlusion problem of clothing key point detection,a bi-directional tree structure model is proposed to make full use of the rich prior structure information of clothing.According to the structure of clothing itself,each keypoint of clothing is connected by a tree structure,and the information flow in the tree structure is transmitted by convolution,so that the keypoints are associated.The core idea of this model is to use clothing structure information to construct the association between the keypoints in the keypoint detection,carry out propagation and interaction,so as to be able to accurately predict the keypoints that are occluded.At the same time,combining high-resolution shallow features and deep semantic features,introducing features of different scales for learning to achieve more accurate detection.(2)In order to alleviate the influence of factors such as the complex background of the clothing image and the wide variety of clothing,a clothing attribute classification method based on the hybrid attention model is proposed.This model combines three attention models:clothing priori,channel and space attention,self attention.Clothing priori mainly uses clothing keypoint information to guide attribute classification.Channel attention is mainly to learn the weight of feature channel.Spatial attention is to learn the spatial position information of clothing.Self attention captures global feature information by modeling long-distance dependence.Finally,the features obtained by the three attention models are further fused and learned.The attention module makes the learning of the model concentrate on the clothing area,and discards useless or irrelevant features to obtain a more robust clothing feature expression.This method makes full use of the attention mechanism and deep neural network model,effectively alleviating the problems of complex and diverse clothing image backgrounds and a wide variety of clothing.In this paper,through the research on the keypoint detection and attribute classification in the field of clothing vision,we improve the algorithm to improve the performance of the model,and promote the development of clothing visual analysis related research fields,such as clothing image retrieval,personalized clothing recommendation,virtual fitting,clothing matching and other applications,so as to promote the development of clothing e-commerce.