A parallel distributed algorithm for feature extraction and disparity analysis of computer images

A parallel distributed processing method for extracting features in stereo images is presented. The algorithm is edge-based and employs the epipolar constraint. First edges are detected in a multilayered network using two methods of locating the magnitude of directional first derivatives and the zero-crossing of second derivatives of smoothed images. The combined edge detection process helps to eliminate `phantom´ edges. Then features that are likely to be detectable in both images are selected. These features are edge intervals and edge orientation, that is, the intersection of image edges with the epipolar line and the corresponding slope of the edges at those points. The feature extraction process is implemented in a two-layered parallel distributed processing model. The first layer provides edge interval representations. The second layer computes a similarity of measure for each pair of primitive matches which are then forwarded to the second stage of the algorithm. The purpose of the second stage is to turn the difficult pixel correspondence problem into a constraint satisfaction problem by imposing relational constraints. This constraint satisfaction is then solved using a neural network. The results of computer simulations are presented to demonstrate the effectiveness of the approach