Gabor filter-based computational models for low-level vision
Kameswara Namuduri · 1992
In recent years, it has been shown that Gabor filters resemble the receptive profiles of simple cells in the visual cortex and hence are used to develop effective computational models for a wide spectrum of machine vision applications. In this dissertation, Gabor filters are studied in the light of the new results regarding wavelets that have gained significant importance in mathematics and signal analysis during the past few years. Based on this study, two powerful computational models based on Gabor filters are developed. The first computational model is for simple edge detection. In this model, a design criteria is established through mathematical analysis for effective edge detection and the performance is compared with that of Canny's optimal edge detector and the first derivative of Gaussian. The second model developed is for feature detection at multiple resolutions. The wavelet characteristics of Gabor filters are utilized in deriving two different schemes for computing multiresolution responses of a family of Gabor filters. In the first scheme, a family of Gabor filters are generated at multiple resolutions starting from the filter at the highest level of resolution and proceeding towards the lowest level. At each level, the resolution of the filter is dilated by a convolution operation. In the second scheme, a bottom up approach is followed in which the filters at multiple resolutions are generated starting from the filter at the lowest level of resolution and proceeding towards the highest level. At each level, the resolution of the filter is increased by a subsampling operation. These two schemes are powerful and can be used for any type of feature detection. The computation of multiresolution responses of Gabor filters is a computationally intensive task which limits their use in real time applications. Hence, special purpose hardware architectures are proposed for computing Gabor filter responses at multiple resolutions. These architectures are derived exploiting the symmetry, separability and wavelet characteristics of Gabor filters in order to obtain high speed and throughput.