Block Implementation of Digital Filters

A crucial problem occurs when attempting to filter an image of large size, caused by computational limitations in processing time and storage capacity. In particular, it is possible to obtain computational gain when operating on a large image by an FIR filter whose impulse response size is much smaller than the region of support of the input image. (The region of support is the set of values on which the image is defined.) To achieve this, a technique known as sectioning may be used, whereby the image is broken up into overlapping sections of size of the same order as the impulse response of the filter, and the filtering operation performed on the sections in turn. Breaking the image into blocks in this manner facilitates the implementation of the algorithm on a parallel processor. There are two sectioning procedures, known as the “select-save” and “overlap-add” methods,(1–3) that may be used to eliminate any wrap-around error caused by the circular convolution inherent in the use of FFT algorithms.