Small object counting with cellular neural networks

Abstract

This report presents a completely cellular neural network-based system architecture for small object counting, where the center positions of small patterns of known shape, size and orientation are located in an input image, in order to be finally counted. The system consists of three cascaded image processing stages: preprocessing performs noise filtering and contrast enhancement, pattern matching approximately locates object positions, and isolating ensures uniqueness of perceived object center locations. Some templates for isolating are presented; their stability is proven.<>