Early segmentation in video compression using CNN processors

Abstract

Two analogic (analog and logic) CNN algorithms are presented which segment a video sequence into objects. The algorithms are mainly based on 3 by 3, linear templates. This allows the CNN Universal Machine to execute the task achieving enormous computation speed (10/sup 12/ equivalent operation per second). The proposed segmentation algorithms rely on texture and contour information only. They differ in the use or not of the color information. The estimated execution time proves that the proposed segmentation method may be implemented in real time. This result and the quality of the obtained frame description are very appealing in the context of the new video coding standard MPEG-4.