Based On Edge Perception And Part Relation Reasoning Of Human Parsing

With the continuous development of deep learning,Human Parsing is more and more mature.Human Parsing is very helpful to other research fields,such as: Motion Recognition,Pedestrian re-identification,Virtual Clothing and so on.Its application scenarios are more and more extensive,such as: video surveillance,intelligent dressup,human-computer interaction,etc.Human Parsing identifies and classifies various of a body’s parts or clothing,which is an accurate pixel-level classification in an image.However,there are still some problems that need to be solved urgently in Human Parsing.For examples,complex human background,inaccurate analysis of the edges of human body parts and insufficient utilization of human body structure information.This thesis uses computer vision and Deep Learning to study the three problems.The specific research work is as follows:(1)Aiming at the problem of complex human background,the thesis proposes a Background Separation Network(BSN).The Human Segmentation is introduced into Human Parsing to help distinguish the foreground from the background.In order to make effective use of the multi-layer features of Residual Network,this thesis proposes a Guide Feature Module and a Supplementary Fusion Module to obtain the foreground feature that can distinguish the foreground and background more accurately.They can further improve the performance of Human Parsing network.The ablation experiment and contrast experiment on LIP dataset show that the m AP reaches 65.15% and the m Io U reaches 55.23%,which are 3.95% and 3.12% higher than the baseline network respectively.BSN has a certain improvement compared with other methods.(2)In order to solve the problem of inaccurate edge analysis of human body parts,the thesis proposes Edge Perception Network(EPN).EPN refines the rough results of Human Parsing guided by the edges of human body parts.They optimize and benefit each other.Thus,it can improve the segmentation effect of Human Parsing.In this thesis,an Edge Perception Module is proposed,which uses the middle-level features and high-level features of Res Net to emphasize the features,which along the boundary of human parts.The module capture more accurate edge features and get the edge map of human parts,that can improve the performance of edge branches and improve the accuracy of the model.The m AP and m Io U of EPN network reaches66.22% and 56.79% respectively on the single dataset LIP,which are increase by5.02% and 4.68% respectively compared with the baseline network.On the multiperson dataset CIHP,the m AP reaches 76.29% and m Io U reaches 66.41%.EPN is very competitive.(3)In order to solve the problem that human body structure information is not fully utilized,this thesis proposes a Part Relation Reasoning Network(PRRN).PRRN deeply explores the relationship between different parts of human body,and puts forward a spatial dependence relationship.That is,there is a spatial connection between left arm and left leg.Similarly,there is a spatial connection between right arm and right leg.Based on the information of human body structure,a part relationship model is established,including composition relationship and dependence relationship.Then the features are reasoned on the art relationship model.PRRN conducts ablation experiments and comparative experiments on LIP dataset,and the m AP and m Io U reaches 63.14% and 54.07%,respectively,which are 1.94% and1.96% higher than the baseline network.