Theory And Application Of Human Semantic Segmentation Based On Deep Learning

With the widespread application of mobile devices and the popularization of 5G networks,people are constantly enriching the ways of obtaining video images.Among these visual images,images based on the human body occupy a large proportion,which provide a large of sufficient data for the intelligent task research of the human body.Among them,human semantic segmentation and skeletal motion recognition are the entrances of most human-oriented intelligent visual processing,which provide data support and more possibilities for higher-level tasks such as human behavior analysis.This article has conducted research on the above two tasks.The main contributions are as follows:In this thesis,we design a novel convolution unit,one-dimensional diagonal convolution unit,which contains the left diagonal convolution unit and the right diagonal convolution unit,to extract diagonal features from the input feature map.On human semantic segmentation tasks,this kind of convolutional unit can correct the mistakes occor at the segmentation boundary.Our diagonal convolution unit can better extract the effective information at different semantic boundaries of the image and adapt to the directionality of the image itself.At the same time,the two diagonal convolutional units can adapt to various situations at different semantic boundaries.In view of the defect that the semantic segmentation model has a large number of parameters and is difficult to be applied to mobile devices,a convolution block containing multiple convolution units is designed to reduce the parameter number.Our convolution block,namely directional convolution block,is based on the diagonal convolution unit above,which contains four one-dimensional convolution units and one two-dimensional convolution unit.Directional convolution block can effectively replace the original standard convolution unit,and reduce the parameter number of the model without influencing the final performance of the network.On the task of human semantic segmentation,this paper verifies the effect of the proposed direction convolution block.On the Pascal Person-Part dataset,we achieved a 63.4% miou with the replaced model,which is an increase of 0.7 percentage points compared to the original model,and slightly reduces the number of network parameters.Based on the skeleton sequence motion recognition task,a new type of spatiotemporal convolution unit is designed.The original spatio-temporal graph unit did not make full use of time and space information,and the new spatio-temporal graph convolution unit designed in this paper reorganized the utilization form of time and space information,and divided four branches to obtain time and space information.And designed two ways to merge different information.On the NTU-RGBD dataset,our method achieved the accuracy rates of 93.51% and 86.29% on the Cross-view benchmark and Cross-subject benchmark,respectively.We obtained an improvement of 0.7% and 1.1% compared with the original spatio-temporal graph convolution unit.