Abstract: A serial neighborhood processing system, preferably employing a pipeline of individually programmable stages, is used to detect image correspondence between multiple frames of image data. Similar features in each frame are extracted utilizing a series of dilation and erosion sequences. In particular, landmarks in each frame are condensed to single image points having exactly one corresponding point in the other frame surrounding it within a given mathematical distance. The second frame of pixel data is registered with the first frame as a function of the spatial relationship between these pairs of matched points in the two reduced frames. A comparison is made between the first frame and the registered second frame to detect differences therebetween. Motion detection, stereo projection, registration of multidate Landsat imagery and matching images of electrophoretic gels are examples of particular uses for the techniques disclosed.