A novel algorithm for image segmentation using time dependent interaction probabilities

Abstract

For a consistent analysis of a visual scene the different features of an individual object have to be recognized as belonging together and separated from other objects and the background. Classical algorithms to segment a visual scene have an implicit representation of the image in the connection structure. We propose a new model that uses an image representation in the time domain, operating on stimulus dependent latencies. Such stimulus dependent temporal differences are observed in biological sensory systems. In our system they will be used to define the interaction probability between the different image parts. The gradually changing pattern of active image parts will thereby lead to the assignment of the different labels to different regions which leads to the segmentation of the scene.