Testing the effectiveness of Non-Linear Rectification on gabor energy

Abstract

The gabor family of filters has received a great deal of attention as it seems to closely approximate the type of low-level processing found in biological vision systems. It has also been widely held that further processing stages involving energy calculation and Non-Linear Rectification (NLR) should be employed. Recent performance evaluation tests of gabor-based features in texture classification [1], where NLR was not used, produced poor results and the hypothesis that use of NLR would improve performance was made.

The work presented here explores the extend to which such a hypothesis can be valid. Two sets of gabor energy features are extracted, before and after an NLR operator is applied to a filtered image; these two sets form the initial texture classification feature set. Feature effectiveness is evaluated using a statistical redundancy test and features are ranked based on the resulting Redundancy Factor (RF). Best features are selected among those with the lowest RF values and subsequently used for classification. Thus, the RF-test provides a strong indicator regarding the effectiveness of features with respect to pattern discrimination. Based on that, the relative effectiveness of pre-NLR and post-NLR features is compared and conclusions on the NLR hypothesis are drawn.