Gabor filters and Gaussian derivatives are flexible parameterizable functions that analyse textured patterns in the images. The advantage of Gabor filters over Gaussian derivatives is related to a higher flexibility in the definition of the function shape, because of a more general set of parameters (degrees of freedom).
To compare the performance of Gabor Filters vs Gaussian Derivatives in practise, we have performed a test where an image region is described by a vector of responses of either Gabor Filters or Gaussian Derivatives. Assuming that the vector of measurements follows a normal distribution, we compute the mean and covariance matrices of such responses in a large training set. For recognition, we compare similar patterns in different images, using the Mahalanobis Distance.
IST performed tests in order to: (i) compare the performance between three similar representations (Gaussian derivatives, directional derivatives, and Gabor filters), and (ii) improve the performance of the Gabor features. We use 82 subjects from the AR face database, half of them used for training, and the remaining half for testing (compute the Mahalanobis distance). We look for eyes in face images, considering a positive match when the global minimum in the image of the Mahalanobis distance is located within a circle of radius r around eye's center. Results of comparing Gaussian derivative vs Gabor feature performance are:
|Gaussian derivatives||Directional derivatives||Gabor filter|
Because of the higher flexibility in the choice of parameters, Gabor filters can represent more efficiently several image regions, surpassing the performance of the Gaussian derivatives and Directional derivatives. In related work inside CAVIAR project [Moreno 2005a, Moreno 2005b], we explored the parameter selection of Gabor filters in order to adequately represent several image features.
The following image shows detected left (red) and right (blue) eye
A paper that describes the process is:
Back to CAVIAR home page.