next up previous contents
Next: Stereo Vision Up: Texture Recognition Previous: Getting started   Contents

Practical

Gabor filters are often presented a solution to texture recognition problems in machine vision largely because of support from psychophysical experiments. The 2D Gabor function is a harmonic oscillator, composed from a sinusoidal plane wave at a particular frequency and orientation, within a Gaussian envelope.

Using the fgabor function in the create tool generate Gabor filters. This function generates filters in the Fourier domain so you will need to run ffti to see the spatial equivalent. Adjust the parameters gb_k, gb_b and theta and see how the resulting filter changes.

Load the broadatz.aiff image in the image directory onto the stack. As convolution is essentially a pattern matching process, it is possible to build filters which respond well to particular types of texture structure of fixed scale and orientation. Attempt to extract the different regions in this image using your knowledge of the way in the which the parameters control the kernels to match them to the scale an orientation of the different textures. Create Gabor `tuned' to a type of texture shown in the image. Apply the filter to the image as a multiplication in the Fourier domain, *. Transform the result back into the spatial domain using ffti and calculate the magnitude of the response. This is done by taking the sqr of each component of the complex result (cast to float), summing + and computing the sqrt. You should see some areas where the filter gives better responses than others. Try to assess the sensitivity of the response to parameter change by seeing how much the parameters can be varied before the high response is lost.

Brodatz textures are particularly suited for texture recognition, there are no illumination artifacts and no perspective effects distorting the appearance of the texture. These issues and the problem of unpredicatble responses from the texture filters at boundaries, are the main obstacles to using such filters in general applications. If you have time try using the house.l.aiff image. Can you find a filter tuned to extract the window shutters? What problems might arise using this technique if the object was viewed from a different direction?


next up previous contents
Next: Stereo Vision Up: Texture Recognition Previous: Getting started   Contents
root 2017-11-20