Systolic arrays for 2D decimation filters

Convolution is required by many image processing procedures such as pattern matching, edge detection and resizing. Decimation filtering is a subset of convolution in that the convolution of an image includes decimation filter products. This paper investigates systolic array designs for two-dimensional convolution and designs for dedicated decimation filters. While existing designs for two-dimensional convolution are suitable as decimation filters they can be improved by reducing the number of processing cells. A new systolic array for convolution/decimation is presented which completes in time proportional to the image size with the array proportional to the window size. The array has optimal latency for the algorithm used. It is derived using a synthesis procedure for multidimensional multirate arrays which includes restricting the size of the array and reducing it from three dimensions to just two.