On local linear transform and Gabor filter representation of texture

Presents a comparative study of two approaches to texture representation under the computational framework of multichannel spatial filtering. The two approaches compared are the Gabor multichannel receptive field model and local linear transform techniques. In the local linear transform approach, it is shown that the orthogonal masks derived from the discrete cosine transform (DCT) have some interesting features which resemble Gabor filters. Although the texture representation concepts of orthogonal DCT basis masks and Gabor filters are fundamentally different, it is demonstrated that the representational structure of both approaches is inherently identical. This allows one to devise an experiment to compare the two approaches in terms of their computational efficiency and the representation potential of the extracted texture features. The performance of the approaches is assessed by means of their segmentation result accuracy on a set of images composed of Brodatz textures