Asymmetrical filters for vision chips: a basis for the design of large sets of spatial and spatiotemporal filters

The problem of actual vision machines on VLSI is the implementation of a large set of linear filters selective to features such as edges, corners, orientations, motion. This procedure requires large arrays of filters with complex structures. In order to reduce the complexity, we propose to implement a simple basis of filters able to generate more complex filters such as oriented quadrature band-pass filters, quadrature wedge filters, velocity tuned filters. The current basis consists of asymmetrical filters implemented with cellular neural networks having only one layer and interactions with four neighbors. We show that this basis generates relevant filters for vision applications and show how the particular structure of the filters basis allows the direct implementation of spatio-temporal filters with a low additional cost