EFFICIENT COMPUTATION OF GABOR FILTER BASED MULTIRESOLUTION RESPONSES

Multiresolution representation in machine vision systems provides a means to detect, analyze and interpret the information content of an image at multiple resolutions. The pyramidal representation of images suggested by Burt and Crowley and the hierarchical construction of orientation and velocity tuned filters developed by Fleet and Jepson are some of the examples for multiresolution representation of image details. In recent years, Gabor filter based methods have been suggested by several researchers for machine vision applications such as edge detection, texture classification and optical flow estimation. These applications require generating a family of Gabor filters tuned to several resolutions and orientations and computing their responses. The computation of Gabor filter responses at multiple resolutions and orientations is a very computationally intensive task. Two efficient schemes for computing Gabor filter based multiresolution responses are proposed. The first scheme namely, recursive filtering method generates a family of Gabor filters and their responses starting from a high resolution and proceeding towards the lowest resolution in steps of an octave. The second approach namely, successive sampling method generates the same responses in the reverse order. In both these methods the symmetric, asymmetric and wavelet nature of Gabor filters are exploited in order to speed up the computation. The applicability and usefulness of the schemes are established through experimental results. Special purpose architectures are proposed to implement the proposed techniques in hardware in order to further speed up the computation.