Enhancing Video Object Segmentation Results Through Biologically Inspired Postprocessing

Object segmentation from a video stream is an essential task in video processing and forms the foundation of scene understanding, object-based video encoding (e.g. MPEG4), and various surveillance and 2D-to-pseudo-3D conversion applications. Many segmentation approaches are pixel-based and suffer from noise in the segmentation results, due to the fact that these approaches do not exploit spatial information. Morphological post processing is typically used to enhance the segmentation results. Here, an alternative post-processing approach is presented. The proposed approach is inspired by well-known aspects of the primate visual system function. The approach is particularly suitable for use with the probabilistic segmentation algorithms and allows for efficient neural-network- based implementation.