Accelerated calculation of gabor features in spatial domain

Abstract

A lot of computer vision tasks are tried to solve by using Gabor features that proved to be very effective feature descriptors. One of the main drawbacks of using Gabor features in real-time practical applications is computational heaviness. Convolution with Gabor filters in frequency domain generates responses of Gabor features at each position in the image and is time consuming operation when Gabor feature consists of tens of Gabor filters. In this article, we introduce the concept of calculating Gabor features at regular grid in the image and show how this together with a generalized separability, symmetry and anti-symmetry of Gabor filter can be exploited to calculate Gabor features in spatial domain even faster than in frequency domain. Ill. 6, bibl.20 (in English; abstracts in English, Russian and Lithuanian).