Local motion detection by hierarchical neural network
Eiji Atsumi, Mikio Takagi, Kazuhiko Yokosawa · Systems and Computers in Japan · 1994
Abstract Perception of visual motion is thought to consist of two stages: generation of candidates and their interaction to determine true motion. In this paper, a three‐layered neural network is applied to detect local visual motion. The network, after learning, could categorize nine types of motion and obtained a motion detection algorithm that included the two states of perception. The internal representations for the first stage agree with the functions of one‐center and off‐center cells, and those for the second stage agree with the functions of lateral inhibition. We also tried to detect arbitrary motions by combining multi‐resolution representation of images with the neural network.