摘要： 借鉴静态图像中通过计算像素灰度的谱残差检测显著性区域的方法,提出基于运动谱残差的视频显著性检测算法。通过提取每一帧的运动矢量场,对矢量场的水平和垂直分量分别计算运动谱残差,融合2个方向的运动谱残差形成运动谱残差图,将谱残差图进行顶帽变换形成视频显著图。实验结果表明,该算法能准确地分割出各种运动场景的显著性区域,在检测效果和抗噪能力方面优于现有视频显著性检测算法,适用于运动微生物检测、行人车辆检测等领域。

关键词: 视频显著性, 谱残差, 光流法, 运动矢量, 顶帽变换, 图像增强

Abstract: This paper references the idea of detecting salient area in still image through calculating spectral residual of the gray-level pixel value,so proposes a video saliency detection algorithm based on motion spectral residual.It extracts each frame’s motion vector field,calculates motion spectral residual of the horizontal and vertical components of motion vector field,merges the two components’ motion spectral residual to form a motion spectral residual image,and finally exertes a top-hat transformation on the motion spectral residual image to form video saliency image.Experimental results show that this algorithm can accurately segment saliency region out of various motion scenes,outperform existing video saliency detection algorithms in terms of detection accuracy and noise resistance.The algorithm can be applied to such field as moving microbe detection,pedestrian detection and vehicle detection.

Key words: video saliency, spectral residual, optical flow method, motion vector, top-hat transformation, image enhancement

中图分类号: 

• TP37