A 60x60 Processor Array Smart Sensor for Binary Image Processung

Smart sensors could be useful in a lot of applications for a speedy and rough vision. We describe in this paper an architecture of a SIMD mesh array optical sensor, we give an implementation in this architecture of the main loop of an algorithm of pseudo-euclidian skeletonization and we present an example of control programmation, a counter of binary pixels made with elementary projections. Whereas computer vision research focuses on grey-level picture processing, a large number of industrial applications still use binary pictures [YOTE88]. As a very high processing speed is required, specialized architectures have been studied for a while [Reeves84]. Thanks to VLSI, high performance compact machines are now possible [Person88]. In this paper, we describe a monolithic parallel binary picture processing, usable as a smart sensor.. . 1 n u p. h. Efficient algorihms on binary pictures have been designed for a number of situations : These algorithms can be built out of iterations of a small number of primitive operations :-unary local operator, in the sense of [Klette80], which compute a result picture out of a given picture through local boolean computations ;-pointwise boolean operators between several pictures ;-picture area measurement. We have designed a parallel p r o c e s s o r performing most of these algorithms while fitting in a single chip and including optical sensors. Its primitive operations are pointwise boolean operations and 2-D shifting in four directions (North, East, West, South). These primitive operations, when iterated and cascaded, provide most of the previous algorithms. By following this implementation, local operators of arbitrary size and shape defined by templates including "don't care" sites are easily achievable in a time linear with templates size and number. Some examples of grey level picture processing are achievable with the optical sensor. A description will appear in [GRZ88]. The idea of a cellular machine including optical input through photodiodes comes back to UCPRl by M.Duff [Duff 671. The successive evolutions of CLIP (number 1, 2, 3) led to the well-known CLIP4 array [Duff 861 which is well suited to both logical operations on binary pictures and numerical operations on multilevel pictures. Other cellular arrays have been built for greylevel picture processing, such as DAP [Reddaway 73land MPP. Finally the GAPP can be considered as a buiding block for large mesh arrays. NOW as far as smart sensors are concerned, a monolithic realisation is required. But while …