On the algorithmic complexity of digital image processing filters with winograd calculations

Abstract

This paper analyzes the computational complexity of methods of digital filtering of images by linear spatial filters. A comparison of the direct method of realization of two-dimensional filtering and the method of realization of two-dimensional filters by Winograd is carried out. It is shown that the best result using the Winograd method is obtained for averaging Gaussian filters and for symmetrical filters of a general form with different coefficients. The Winograd method application for Gauss filter reduces the number of multiplication operations 3 times at an increase of several addition operations 1.84 times. For symmetric general filters with different coefficients, multiplication operations are reduced to 2.12 at a rise in the number of addition operations 2 times.