Research On Clothing Key Point Detection Method Based On Pyramid Network

Clothing online shopping has become an important way for modern people to choose to buy clothing.People not only buy clothes conveniently,but also put forward personalized needs for clothing style,fabric and color.In order to describe the characteristics of various clothes as accurately as possible,we can use the key points of clothes to analyze the local details of clothes.The positioning of clothing key points is a basic work,which can promote the application of online fitting,clothing part alignment,clothing local attribute recognition,clothing image automatic editing and so on.It plays a key role in the analysis and application of fashion big data.However,there are many kinds of clothing data,large posture changes,clothing occlusion and other problems.The existing studies mainly judge the location of clothing key points according to clothing texture,which can not well adapt to the flexible and changeable characteristics of clothing,and can not well solve the problem of clothing occlusion.Therefore,the main work of this paper includes:(1)Due to the problems of many clothing categories,complex posture and small-scale occlusion,the traditional clothing key point detection method leads to the loss of some key point information and is difficult to establish the relationship between key points.To solve these problems,a two-stage clothing key point detection method based on CPM and affinity vector is proposed.Firstly,the target detection method is used to reduce the impact of clothing category and background on clothing key point detection.Secondly,the human body key point detection model CPM is used and improved to detect clothing key points,The affinity vector is used to constrain the coordinate information of adjacent key points,enhance the relative position relationship of adjacent key points,and use the spatial constraints between key points to improve the accuracy of detection.Finally,a two-stage garment key point detection framework is implemented.Experiments show that the framework can detect clothing key points more accurately,and has strong robustness to clothing key points dispersion,multi category and small-scale occlusion.(2)Because the factors such as clothing occlusion and multi angle have an important impact on the accuracy of clothing key point detection,the ability of neural network to extract clothing key point features is limited by clothing occlusion and key point information fuzziness.This paper proposes an occlusion aware clothing key point detection method based on pyramid network and human posture.Firstly,the autocorrelation of key points is expressed through the attention mechanism to alleviate the problem that it is difficult to capture the partially occluded clothing texture features,and the discrete points are prevented from being missed according to the sparsity analysis of key point eigenvalues.Secondly,through the selection of effective highlevel features of key points,redundant calculation is avoided,and the correlation calculation is carried out with the help of the similarity between human joint points and clothing key points,so as to alleviate the semantic fuzziness of key points in occlusion area.Experiments show that this method has a certain spatial understanding ability,has a strong ability to perceive the occluded objects in clothing,and can effectively make up for the semantic fuzziness of occluded clothing.