Human Parsing Based On Deep Learning

Human parsing has now become a relatively popular research direction,and the technology has many application scenarios,such as video monitoring,smart home,and human-computer interaction.Human body analysis and semantic segmentation are essentially the same,and belong to the classification problem at the pixel level,so most of the human body analysis models are improved based on the semantic segmentation model.However,many research work ignores the problems of the structured prior information between each part of the human body and the correlation between each pixel of the human body picture,and the experimental results of many studies show that the human body analysis performance is poor in complex scenes where multiple people exist,and it is impossible to outline all individuals in the predicted pictures.For the above problems,this article uses the deep learning method to analyze the human body in the natural state from the perspective of single-person analysis and multi-person analysis.The main work is as follows.Firstly,based on the research status of semantic segmentation and human parsing,the relationship and difference between semantic segmentation model and human parsing model are analyzed,and the problems existing in humam parsing model based on semantic segmentation are summarized and improved methods are proposed.Secondly,3533 human pictures were extracted from PASCAL VOC 2012,and the human tags were fused,and the PASCAL-Person-Parts-5 dataset and PASCAL-Person-Parts-7 dataset were established.The PASCAL-Person-Part-Edges-7 dataset was made using the Eight-Neighborhood edge point extraction algorithm.Thirdly,from the perspective of single-person analysis,it is found that there is structured prior information between various parts of the human body picture,and compared with scene pictures,there is a stronger correlation between the pixels of the human picture,so a human parsing model based on the conditional random fields and edge point penalty mechanism is proposed.Using conditional random fields to consider the influence of adjacent pixels and optimize the prediction results,and the edge point penalty mechanism is used to model the structured prior information of each part of the body.Experiments on the PASCAL-Person-Parts-5 and the Freiburg sitting people datasets prove that the model can improve the performance of human parsing model and improve the outline effect at the edges.Finally,from the perspective of multi-person analysis in complex scenes,it is found that the human body analysis model has a poor analysis effect on complex scenes of multiple people and connot accurately locate all individuals in the pictures.Therefore,in order to improve the positioning process of human parsing model,a human parsing model based on edge detection is proposed.Experiments on the PASCAL-Person-Parts-7 and PASCAL-Person-Parts-Edge-7 datasets prove that the model can indeed strengthen the positioning process,improve the analytical performance of the model,and improve the effect of multiple human parsing.