Networks for image analysis: motion and texture

This study outlines certain difficult problems in image processing and perception that can be solved by simple cooperative and competitive neural interactions. A scheme for measuring local image velocity in a way that is disambiguated from the form of the moving object and from the aperture through which it is observed is discussed. A basic theorem is introduced, called the spectral coplanarity theorem, which shows how cooperation and competition in a winner-take-all network allows the outputs of neurophysiologically measured spatiotemporal filters to be combined to yield velocity information independent of the form of a moving object or its contour. The classical concept of a neural receptive field is generalized to the concept of a velocity receptive field, in which inputs are summated along an inclined plane in a 3-D space of linear filters that are tuned (as striate cortical neurons are) to orientation, spatial frequency, and temporal frequency.<>