A Two-dimensional Systolic Array Processor for Image Processing

Abstract

processing and since-an 8-bit multiplier -is expensive in terms of silicon area, most previously proposed twodimensional arrays use a single-bit processing element. levels bit processor, but can yield significant improvements in The performance of sequential computers in the rapidly expanding fields of image processing and vision has been found to be quite inadequate since most From our analysis, we have found that a 4-bit processing applicationsProcessing Of the image frame(s) in element requires a slightly kgFer silicon area than a time' A ' 12 512 image frame with 256 requires roughly 256K pixels of 8 bits each for storage performanm. and most applications today, e.g. optical flow and motion This paper reports on our on-going research towards analysis, require the storage and processing of mutliple vLsl processor architecture for ima,ge processing applications.processing, the obvious solution is to design parallel The processor architecture is based on a sIMD array of 4-algorithms and architectures that can be bit processing elements, interconnected by a mesh image network with four nearest neighbors. Each processing processing and vision techniques are based on spatial frames at high speeds' with the rapid gowth in design and ~p l e m e n~t i o n , of a two-dimensional array and the advances made in vLsl and wsl Many 0-8186-2465-5IQ1 $1.00 0 1991 IEEE 215

SMAC: A VLSI Architecture for Scene Matching

“…The mesh architectures are found to be well suited to low-level image processing tasks where the (2D) image is mapped on to the processor network topology. Examples of the mesh systems are the SliM [10], the MGAP [11], the SIBA [12], the GRID [13], the LIPP [14], the PCLIP [15], the MPP [16], the GAPP [17], the BLITZEN [18], and the CLIP [19] series. These systems have been developed primarily for low-level image processing applications.…”

A linear array processor with dynamic frequency clocking for image processing applications

“…In the past, several special-purpose image processing machines such as SIBA, RVS-2, and MGAP have been proposed and built [18], [6], [10], [8]. These machines differ from each other in their architecture, interconnection, granularity, and technology.…”