一种仿生物视觉感知的视频轮廓检测方法

Dynamic Contour Detection Inspired by Biological Visual Perception

消除背景的局部边缘干扰同时保证目标的完整轮廓是视频轮廓检测的一个难点,基于运动感知的生物视觉证据,提出一种运动能量抑制模型,有效抑制背景边缘,激励目标的强边缘.通过归一化整理视频运动切片的四方向运动能量抑制响应,反映V1层视觉神经元的周围抑制感知特性,进而采用"双半圆盘"算子提取边缘梯度响应,同时,结合运动和外观线索,用随机森林对边缘梯度响应的局部结构进行树划分,得到最终的检测结果.实验表明,本文提出的方法较已有的视频轮廓检测方法有更优的量化查全–查准率曲线、F-measure值和AP值以及更好的视觉轮廓感官效果.

There is a primal challenge to eliminate local edges from noisy clutter while simultaneously preserving the complete object silhouette in dynamic contour detection. Inspired by biological evidences for visual motion perception, we formulate the motion energy inhibition model as a computational mechanism for effective background suppression and foreground enhancement in boundary responses. The normalized integration with four-direction-channel motion-filter response in spatio-temporal slices reflects the dynamical ＂surrounding-suppression＂ characteristic in V1 visual neuron, which uses two half-disc structure to extract contour gradient. Finally, we exploit the random forest model to partition the contour gradient from jointly motion and appearance cues in tree-like style to achieve object contours in video. Experimental results demonstrate better performances of this approach in quantitative precision-recall curve, F-measure and AP values, and qualitative visual effects.