A fast image filtering processor using the fermat number transform

This paper discusses a newly developed high‐speed image filtering processor, aiming at high‐accuracy and high‐speed isotropic filtering. The isotropic filtering is made efficient by the following features. (1) The image is sampled over the triangular lattice and processed by the hexagonal finite‐impulse response (FIR) filter. (2) The filter is realized by the Fermat number transform (FNT) with hexagonal sectioning procedure. The high‐speed hardware has the following features: (1) FNT‐butterfly circuit using 3‐input adders; and (2) multiple processing by the memory structure with three‐level hierarchies. By those features, the proposed system executes the filtering in approximately 0.5 s for an image composed of 512 × 512 pixels in triangular lattice (512 × 591 pixels in square lattice) by a hexagonal FIR filter with edge distance of 15 (17 × 17 in square lattice‐square FIR filter). This paper describes the features, architecture and operation of the proposed system. Then the result of correcting the isotropic blur is presented, indicating that a practically sufficient accuracy is obtained.