时序特征引导的视频目标分割方法OA
Temporal feature-guided method for video object segmentation
为了解决现有目标分割方法对光流图的过度依赖问题,并充分利用视频目标特有的时间相关性和空间相关性,提出一种先利用目标与背景的时序特征信息提取感兴趣区域,后利用空间显著性特征分割目标的两阶段方法.首先,设计光流异常区域检测器,以序列图像间像素位置变化产生的光流场作为时序特征,学习目标与背景在时序特征上的差异,从而检测出目标位置,并将分割范围引导至目标所在感兴趣区域(region of interest,ROI).然后,使用提出的DC-U2Net作为显著目标检测器分割出ROI中的显著目标,得到掩码后,完成目标分割.其中,DC-U2Net通过引入双通道RSU模块与残差连接路径,进一步提高目标掩码的精度.该方法在两个公共数据集上进行了性能测试,取得了有竞争力的结果.在DAVIS2016数据集上,J分数、F分数与G分数分别达到83.7、86.7、85.2.结果表明,针对视频目标时序特征和显著性特征的特点,在不影响实时性的情况下提出的基于时序特征引导的两阶段视频目标分割方法,提升了目标分割精确度.
To address the excessive reliance of existing object segmentation methods on optical flow maps and fully leverage the temporal and spatial correlations inherent in video objects,a two-stage method was proposed that first extracts regions of interest(ROI)using the temporal feature information of objects and backgrounds,and then segments objects using spatial saliency features.Firstly,an optical flow anomaly region detector was designed,which used the optical flow field generated by pixel position changes between sequential images as temporal features to learn the differences in temporal features between objects and backgrounds,thereby detecting object positions and guiding the segmentation scope to the ROI where objects were located.Then,the proposed DC-U2 Net was used as a salient object detector to segment salient objects in the ROI,obtain masks,and complete object segmentation.DC-U2 Net further improved the accuracy of object masks by introducing a dual-channel RSU module and residual connection paths.The proposed method was tested on two public datasets and achieved competitive results.On the DAVIS2016 dataset,the J-score,F-score,and G-score reached 83.7,86.7,and 85.2 respectively.Aiming at the characteristics of temporal features and saliency features of video objects,the proposed two-stage video object segmentation method guided by temporal features-guided improves object segmentation accuracy without compromising real-time performance.
朱琳琳;李家勋;刘珈溪;梁宵
沈阳航空航天大学 自动化学院,沈阳 110136沈阳航空航天大学 自动化学院,沈阳 110136沈阳航空航天大学 自动化学院,沈阳 110136沈阳航空航天大学 自动化学院,沈阳 110136
信息技术与安全科学
时序特征视频目标分割深度学习掩码两阶段方法
temporal featurevideo object segmentationdeep learningmaskstwo-stage method
《沈阳航空航天大学学报》 2026 (4)
88-96,9
国家自然科学基金(项目编号:61973222)辽宁省高校基本科研业务费专项资金资助项目(项目编号:LJ232410143058).
评论