In this study, Self-Orgarnized neural-network is used to solve the correspondence problem for edge points extracted from a pair of axial stereo image. Edge points are extracted by canny edge operator and matchable candidate points of front image with respect to the rear edge point are searched by scene radience constraints. Edge points of rear image are the output nodes of Neural network and the edge points coordinate of front image is fed into the 2-D input node of Neural network. The weights between output nodes and 2-D input nodes are updated for the accurate correspondence. After several iteration of updating, the weights are converged to the correspondence points of front edge image. Because of the feature map properties of Self-Orgarnized Neural network, noise-free, smoothed depth data can be achieved.
