基于Mask R-CNN和CEDN的人体实例分割方法  

Body instance segmentation method based on Mask R-CNN and CEDN

作　　者：姚砺[1] 高傲 张乃港 吴昊然 万燕[1] YAO Li;GAO Ao;ZHANG Naigang;WU Haoran;WAN Yan(College of Computer Science and Technology,Donghua University,Shanghai 201600,China)

机构地区：[1]东华大学计算机科学与技术学院,上海201600

出　　处：《智能计算机与应用》2023年第9期9-12,共4页Intelligent Computer and Applications

摘　　要：针对人体实例分割任务中Mask R-CNN方法对于边缘分割较差的问题,本文提出了一种结合轮廓检测算法CEDN进行改进的方法。首先,通过Mask R-CNN检测出背景中的人体实例掩码,通过区域填充获取到精细化人体实例分割结果的方法。首先,通过Mask R-CNN检测出背景中的人体实例掩码;其次,利用轮廓检测算法得到精细化人体轮廓,再通过改进的区域填充算法填充出人体分割掩码,从而提高人体分割精度。在LSP数据集上进行验证,本算法相较Mask R-CNN准确率提高了4%,召回率提高了8%。算法有效的改进了Mask R-CNN的分割结果,改善了人体分割边缘较差的问题,进一步提升了人体分割的精度。Aiming at the poor edge segmentation of Mask R-CNN method in the human instance segmentation task,a CEDN method was proposed to obtain human contour and then obtain refined human instance segmentation results through region filling.Firstly,Mask R-CNN was used to detect the Mask of the human body in the background,then the refined human contour was obtained by the contour detection algorithm,and then the human body segmentation Mask was filled by the improved region filling algorithm to improve the accuracy of human body segmentation.Compared with Mask R-CNN,the accuracy and recall rate of this algorithm are improved by 4%and 8%respectively.Experiments show that the algorithm can effectively improve Mask R-CNN segmentation results,optimize the problem of the poor edge of human segmentation,and further improve the accuracy of human segmentation.

关 键 词：人体实例分割 Mask R-CNN CEDN 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]