融合视网膜运动感知的视频目标分割  

RETINA MOTION ESTIMATOR-AIDED VIDEO OBJECT SEGMENTATION

作　　者：吴婷 李椋 王刚 高晋 陈明松 王以政 Wu Ting;Li Liang;Wang Gang;Gao Jin;Chen Mingsong;Wang Yizheng(University of South China,Hengyang 421001,Hunan,China;Institute of Military Cognition and Brain Sciences,Beijing 100850,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]南华大学,湖南衡阳421001 [2]军事认知与脑科学研究所,北京100850 [3]中国科学院自动化研究所,北京100190

出　　处：《计算机应用与软件》2024年第8期240-246,302,共8页Computer Applications and Software

基　　金：科技创新2030—“新一代人工智能”重大项目(2018AAA0102800,2018AAA0102802,2018AAA0102803);国家自然科学基金面上项目(61972394)。

摘　　要：视频目标分割容易受到目标快速运动、目标遮挡等情形的影响,因此高精度的视频目标分割是极具挑战性的任务。利用光流可以促进物体的分割,但运动边界附近区域的光流往往计算得不准确,从而间接影响了基于光流的视频目标分割性能的提升。为突破上述局限,结合生物视网膜大细胞通路模型所提取的运动轮廓信息,来辅助计算运动边界区域的光流,并与传统视频目标分割方法前景-背景分割结合,交替更新光流和分割。所提方法在公开数据集DAVIS-2016、SegTrack-v2、YouTube-Objects上的实验结果表明,该方法相比于基线方法其平均分割精度分别提升了2.2百分点、1.3百分点、1.9百分点。Video object segmentation(VOS)is easily affected by the object fast motion,extrinsic occlusion,etc.,so the high-precision VOS remains challenging in the area.Optical flow can be used to improve object segmentation,but the precise estimation of optical flow around the moving boundary is hard to be fulfilled,which affects the unveiling of VOS performance based on optical flow estimation.To overcome the above limitations,this paper combined the motion contour information extracted by the retinal macro-cell pathway model to assist in calculating the optical flow of the motion boundary region,and combined with the traditional VOS method foreground-background segmentation to alternately update the optical flow and segmentation.Experimental results on several public benchmarks,i.e.,DAVIS-2016,SgeTrack-v2 and YouTube-Objects,show that the proposed method improves the average segmentation accuracy by 2.2,1.3 and 1.9 percentage points respectively,compared with the baseline method.

关 键 词：视频目标分割 视网膜模型 图割模型 光流算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]