Efficient computation of Gabor features

Gabor filter responses are widely and successfully used as general purpose features in many computer vision tasks, such as in texture segmentation, face detection and recognition, and iris recognition. In a typical feature construction the Gabor filters are utilized via multi-resolution structure, consisting of filters tuned to several different frequencies and orientations. The multi-resolution structure relates the Gabor features to wavelets, but the main difference, non-orthogonality, also is connected to the main weakness of the Gabor features: computational heaviness. The computational complexity prevents their use in many real-time or near real-time tasks, such as in object tracking. Fortunately, many arithmetic tricks exist which can be employed to significantly improve the computational complexity with negligible loss in accuracy. The main contribution of this study is the comprehensive survey of existing and development of new improvements which can be applied to filter parameter selection, filter construction, and feature computation. They are combined to provide a complete framework for optimally efficient computation of Gabor features. To make the proposed framework the most valuable and useful the implementation is distributed as public software.