Dynamic Occlusion Contour Detection Inspired By Biological Visual Perception

Contour detection is an important component of low-level visual perception, which is widely used in object segmentation, object recognition, surface reconstruction and other visual tasks. There are two points for being difficulty to extract video contour or boundary in different dynamic scene. Firstly, background clutter can causes unexpected local edge, which may not belong to any foreground objects. Secondly, There are more than one visual cue can estimate local variation to perceive edge. Different cues, such as motion, brightness, color, texture, disparity, depth, can generate inconsistent contour result. In recent years, contour detection biologically inspired by visual perception has been widely recognized and applied. Using the mechanism of neurological visual process is not only improved the performance of contour detection, but also consistent with the results of human visual perception.However, the result of contour perception perceived is difficult to guarantee their occlusion contour. Image segmentation by similarity of pixel appearance marked form region blocks which is complementary with contour, and get the complete region boundaries to ensure that the contour is closed. We give an improved normalized cut method, to transform the result of contour perception into accurate region marked result, resulting in a complete closed contour. Therefore, this paper based on biological contour perception research the occlusion contour and region segmentation method to improve the integrity of the object contour detection.This thesis includes the following four aspects:(1) In order to decrease the false alarm contour detection from background clutter, motion energy inhibition model simulated by biological visual perception is proposed to optimize single-cue for contour detection, and gives reliable explanation for the dynamical’surrounding-suppression’ characteristic in V1 visual neuron.(2) On the basis of motion cue, appearance cues, such as brightness, color and texture are simultaneously combined into edge feature vector. The importance of multi-cues is optimized to generate unified edge response. Due to the diversity of random forest method, effective confidence map for edge response can be generated for each video key frame.(3) For solving the bottleneck problem in high-dimensional feature clustering, down-sampled normalized cut is used to improve normalized-cut method. For reducing the time-consuming of calculation, global inference for contour detection is also discussed in segmentation process.(4) Region merging in multi-scale image segmentation need reliable boundary detection result. Hierarchy alignment is handled for removing unstable edge. Ultrametric contour map can be obtained in multi-scale feature space, which can give accurate multilayer segmentation, eventually form occlusion contour.