Research On The Analysis Method Of The Dressed Human Body For Two-dimensional Virtual Fitting

Try-on can effectively enhance the user's online shopping experience.However,the image of dressed human has complex backgrounds,abundant costumes with rich textures,and different human postures.Therefore,it's important to accurately understand the human centered semantic region in the image for try-on.To address the problem of in accurate human parsing results due to the factors such as human pose,edge contour,complexity of clothing and accessories as well as occlusions of human pose joints in dressing scene,coarse parsing feature,edge contour and pose features were integrated,and more accurate parsing results were obtained by defining the combined function of structure loss function and human parsing loss function.The experimental results show that the proposed model improves the human parsing accuracy and achieves more accurate segmentation results.Because different types,various styles and textures of clothing in the dressing scene affect the positioning of the contour.Moreover,background elements interfere with edge contour recognition,and make it difficult to learn edge features.In this paper,based on Res Net-101,we construct an edge extraction network.It combines the underlying features with global information and high-dimensional features with local information.Our network can improve the accuracy of human body contours and reduce problems such as edge confusion and loss of details caused by the difficulty of accurate positioning of each semantic edge.Then,different poses,overlapped body parts and easily confused joint points lead to the low accuracy of joint point positioning.This paper constructs the dressed pose models.Each of them clusters the human joint points with clustering algorithms to get the posture module.Next,this paper divides the dressed poses into four kinds,and strengthens the network model's ability to locate the human joint points with a proposed dressed posture loss function in the posture estimation network.So we can further improve the accuracy of posture estimation.Finally,there are too many clothing semantics in the image of dressed human,which often cover important human parts.A single feature cannot describe edge information and human part information at the same time.All of those lead to the failure to accurately identify human parts or the boundaries among various semantics and the confusion and wrong recognition in human body analysis results.This paper uses the edge information provided by the edge contour,the joint information provided by the posture feature,and the rough analytic features extracted from the basic network for accurate analysis.What's more we define the structure loss function to obtain more accurate analytical results.