Scale-Adaptive Texture Classification

Scale invariant texture analysis is a fundamental challenge in image processing. As a consequence of the scale invariance, these kind of features are often characterized by a lower discriminative power. We observed, that scale invariant features did not pose a benefit in classification scenarios with varying scales in the training set. This is supposed to be an effect caused by an implicit scale selection done by the classification method. In this work, we analyze this effect based on the k-nearest neighbor classifier. Inspired by this effect, we employ global scale estimation algorithm utilizing scale-normalized Laplacian of Gaussian extrema in scale space, to improve the classification accuracies of scale variant features in a scenario with varying scales. We propose a general framework for scale-adaptive classification, which proved to improve the classification accuracies with a variety of feature extraction methods in such a scenario.