Human Instance Segmentation Based On Convolutional Neural Networks

The main task of instance segmentation is to separate and mark out several categories in the image.Based on convolution neural network,to improve the model.The current instance segmentation algorithm is unable to segment two highly overlapping objects,and previous instance segmentation algorithms basically detect objects in images,and then segment them on the basis of generated detection boxes,which may lead to two very close people being unable to distinguish.This paper found that the particularity of human bones can be used to better accurately segment the people in the picture,thus avoiding the missed detection caused by candidate box.In addition,this paper proposed two new network structures,which greatly improved the accuracy of the segmentation of human instances.The core ideas of this paper are as follows:1.A newly designed convolutional neural network is proposed to improve the segmentation quality by using the human body instance segmentation network based on human posture and scoring mechanism.A new scoring mechanism module is added to describe the segmentation quality of human body more accurately by predicting the IOU relationship between Mask and ground truth.In addition,a network that can directly use and learn IOU is proposed,which can improve the quality of instance segmentation.2.A new network structure is proposed,which is improved by using convolutional neural network which combines human body posture and instance position.The core of the network model is to interpret human body segmentation into two instance objects to complete the category perception prediction task simultaneously.Therefore,this paper decided to divide the experimental data objects into a regular grid,which was defined as S×S.Assuming that the center of the object to be partitioned is in the grid defined,then the grid not only completes the prediction of the instance type,but also divides the object into instances.The network framework in this paper is more accurate and more robust to the results of human segmentation.In general,the method proposed in this paper has been verified by a large number of experiments and compared with other algorithms.It can be seen that the new network model proposed in this paper for human body instance segmentation has good accuracy,and the precision of segmentation has also been improved.In order to obtain more abundant feature information,this paper adopts Res Net and FPN network to extract feature,and integrates multi-layer feature information to make the segmentation result more accurate.The experimental results show that the network framework in this paper is more accurate and more robust for human segmentation...