Modular Arclutectures for Two-Dimensional Digital Signal Processing

This paper addresses the implementation aspects of highly modular two-dimensional (2-D) digital filters. The idea of matrix decomposition 111 is used to provide increased parallelism and regularity. Four different structures namely the transversal, the distributed arithmetic, the stored product and the systolic array, are considered. For comparison purposes, the direct implementation is included. Analysis is performed and comparisons are made in terms of hardware cost, cycle time, finite register length effects and regularity. From the analysis, it is found that the systolic array structure seems to offer the best compromise among the various conflicting figures of merit. The distributed arithmetic structure is shown to be superior in the case of low-order filters, while the stored product structure is preferable for high-order filters. The inherent parallelism and high throughput rate of these structures make them very suitable for real-time image processing applications.