基于空间感知与细化残差的视频运动目标分割

Video moving object segmentation based on spatial⁃aware and refinement residual

针对视频运动目标分割中目标快速移动、障碍物遮挡、相似干扰时存在剧烈位移或目标消失等问题,提出一种基于空间感知与细化残差的视频运动目标分割方法.首先,将标签图像输入小型神经网络获取视觉信息;其次,通过空间感知器对当前帧获得预测的目标热图,以获取空间位置信息;最后,利用空间信息和视觉信息,融合细化残差模块细致轮廓边缘,完成目标分割任务.引入空间感知器,可以减少空间位置的误差,提升了空间调制器网络的准确性;利用多级细化残差模块改善特征图的边缘细节,可以提高目标分割的准确度.在DAVIS 2016和DAVIS 2017数据集上的实验结果表明,该方法可以获得良好的分割结果.

Aiming at the problems of fast moving object in video moving object segmentation,severe displacement or disappearance of objects when obstructed by obstacles,this paper proposes a video moving object segmentation method based on spatial perception and refined residuals.Firstly,we input the label image into a small neural network to obtain visual information.Secondly,we obtain a predicted object heat map for the current frame through a spatial perceptron to obtain spatial position information.Finally,we use spatial information and visual information,and fuse refinement residual module to refine the contour edges to complete the object segmentation task.This paper introduces a spatial perceptron to reduce the error of the spatial position and improve the accuracy of the spatial modulator network.The multi⁃level refinement residual module improves the edge details of the feature map,thereby improve the accuracy of target segmentation.Experimental results on the DAVIS 2016 and DAVIS 2017 datasets show that the proposed method can obtain good segmentation results.