Image Segmentation by a Network of Cortical Macrocolumns with Learned Connection Weights

摘要

Image understanding in the brain or a computer requires segmentation of observed images, i.e., their partition into different semantically-connected parts that each constitute one physical object. This task is fundamental for further processing and analysis of visual information and seems to be accomplished by the brain very easily. Nevertheless it is a very demanding challenge for computer algorithms. In this article, we present a network of neuronal macrocolumns, which processes contour information by favoring closed contours. The connecting weights have been learned from real image sequences before. Then, segmentation is achieved on the basis of color, texture, and contour information.