Gabor filter

Generates Gabor filters using bandwidth, aspect ratio, phase, wavelength and angle as parameters.

Description

#5237 implements a Gabor filter for image processing. From the comments and through my own use, the implementation seems to have limited freedom, and also the variances are defined in the original x,y instead of the rotated x', y' directions.

The code is a modified version based on http://en.wikipedia.org/wiki/Gabor_filter following the conventions and parameters used in http://matlabserver.cs.rug.nl/edgedetectionweb/web/edgedetection_params.html.

The function accepts scalar inputs only, and each call generates one filter. The filter generated is always real. Should you wish to build a complex Gabor filter, you may use (0 phase) + j(pi/2 phase). Filtering can be performed using conv2( ) or imfilter( ).

An example file is included, which mimics the behavior of http://matlabserver.cs.rug.nl/edgedetectionweb/web/index.html at default settings with Lena.jpg image as input. Note that the scaling of numbers is different from the web-app, but consistent with the formulae in http://matlabserver.cs.rug.nl/edgedetectionweb/web/edgedetection_params.html

Update 18-Nov-2010: Thanks to Rachel (see comment below), I have corrected a mistake in the example file. The results are unchanged though. Also, the cs.rug.nl website seems offline.

Acknowledgements

The author wishes to acknowledge the following in the creation of this submission:
2D Gabor Filter(Ver1,2,3)

• Required Products: Image Processing Toolbox
• MATLAB release: MATLAB 7.4 (R2007a)

[David Conger]

 

[Mario]

Please, someone can add some usage example of this code to an image to reproduce "exactly" the same results of the matlab's website http://matlabserver.cs.rug.nl ?

[James Kelly]

hey, do you have any idea how to convert this to c#? I want to convert it but have no idea on how this loops trough pixels on the source image and how it produces the output image...

If you can write c#/c++/java/c code or at least explain to me as in what happens to the pixels that would be great, thanks. :)

[Jon Hornstein]

I superficially looked at the code and this looks like a gaussian filter with polar rotation of coordinates only 1/2 of the answer. Can you please advise.

My thoughts were:
2D gaussian = g(x,y) = 1/(2*PI*Theta(x)*Theta(y)) * exp(-0.5*[(x/theta(x)).^2 +(y/theta(y)).^2])

gabor = h(x, y) = g(x, y)* exp[j*2*pi*(U*x+V*y)]

[Jon Hornstein]

Thanks I see now you are using just the real component

[Rachel]

Very good implementation. And good job normalizing it! What I do notice is: you seem to do (a+b).^2 to get the modulus of the result. I thought for a complex number a+bj, you are supposed to do a.^2+b.^2 for the modulus. Lucky enough for this case, the cross term 2ab is simply zero because we are summing some trigonometric functions from 0 to 2pi.

[Marc]

Can someone explain me why when a bandwidth >1 is used the result differs from http://matlabserver.cs.rug.nl

using a bandwith of 10 I had to replace

    img_out_disp = sum(abs(img_out).^2, 3).^0.5;
with
    img_out_disp = sum(imag(img_out).^2, 3).^0.5;

in order to have the same results as http://matlabserver.cs.rug.nl

but don't know why. I believe this come from the ratio sigma to lamda < 0.5. Any mathematical explanation?

[shuai]

 

[Jason]

Great one, thanks a lot. By the way, may I know, what does the size denote? Also, I can't seem to understand how the meshgrid in your code works.

[Rakesh Chalasani]