Description

Read in the input image, transform it by a DFT and apply a set of one channel filters to this image. The set of filters includes:

• Filtering by a Gabor filter with frequency and orientation selective properties
• Some kind of linear transform on the image pixels. Currently implemented is the absolute value nonlinearity and the tanh() nonlinearity as described by Farrokhnia.
• Averaging (blurring) of the filtered image by a Gaussian lowpass filter

The output from this program is a image sequence where the input image has one output image per filter in the filter set. The output is scaled to the maximum dynamic range of the output format. Be aware that this scaling implies that training a classifier on the output of the filter-response of one specific image and classification on the output of another is impossible, because the two outputs will probably be scaled differently.

See: ``Multi-channel filtering techniques for texture segmentation and surface quality inspection'', Farshid Farrokhnia, Ph.D. dissertation, Michigan State University, Dept. of Elect. Eng., 1990

And: ``Determining Gabor-filter parameters for texture segmentation'', Dennis F. Dunn, William E. Higgins and Joseph Wakeley, SPIE Vol. 1826, Intelligent robots and computer vision XI (1992), Pages 51-63