Research On Human Pose Estimation Methods In Fashionable Dressing Scenes

The human pose estimation in the dressing scene can effectively improve the authenticity and dynamic display of the virtual try-on.Therefore,the accurate result of human pose estimation in the dressed image plays a vital part in assisting the twodimensional virtual try-on.This article aims at the low accuracy of human pose estimation due to various clothing styles,background interference,and various body postures in the fashion dressing scene.To solve this problem,we take the fashion street image as an example.Firstly,to construct a fashion dressing image dataset,we crawl the large-scale website of fashion street image which named Chictopia and combine with collecting online image to obtain the original images of the dataset,and using Label Me to label the images with fine human body information.Then,the pose category loss function in the fashion dressing scene is defined through pose clustering,then construct a pose representation model by adding multi-scale loss and feature fusion based on the stacked hourglass network.Finally,the dress part segmentation is combined to constrain the position of human body joints,and get more accurate human pose estimation results by optimization.Experimental results show that this method can effectively improve the accuracy of human pose estimation.First of all,aiming at the problem of the low accuracy of human detection in fashion dressing scene,a dressed human detection model based on the improved twostage detector is proposed.Firstly,to improve the effect of dressed human body feature extraction,this method improves the feature extraction network of the two-stage detector which named Faster R-CNN algorithm.Then,resetting the default anchor size to suit the ratio of dressed human detection box,and use the region proposal network to generate a preliminary region proposal box containing the foreground.Finally,The ROIAlign is used in the ROI pooling layer to obtain a fixed-size feature map,and input it to the full connection layer to get accurate human detection results in the dressing scene.Secondly,aiming at the difficulty of extracting the feature of human body joints result from the complex spatial background of dressed human,the variety of clothing textures and colors,and the difficulty of the network learning of human pose estimation result from the changeable viewing angle of dressed human,a human pose representation model in fashion dressing scene is proposed.1)To improve the robust of human pose representation model in changeable pose of dressed human and human pose estimation from different perspectives,we define the pose category loss function based on the pose template obtained by clustering dressed human pose in fashion dressing image data set,then combine it and the euclidean distance loss function to construct the multi-scale loss of the pose representation model.2)To learn the local and global feature of human body joints,we construct a dressed pose representation network by adding multi-scale loss and feature fusion based on the stacked hourglass network,which solve the interference of diverse background and the human wearing clothing on the feature extraction of joint points and enhance the positioning accuracy of human body joints.Finally,aiming at the interference of visibility of dressed human pose by the human body wearing clothing,which leads to the mislocation of the joint points in the estimation of human pose,a human pose estimation network combined with the dress part segmentation is proposed.We firstly construct a dress part segmentation network by fusing the deep and shallow features of the residual network.Then use the dressing part segmentation information to constrain the positioning of human body joints.Finally,improving the positioning accuracy of joint points in the estimation of human pose through pose optimization.