Degradation adaptive texture classification

Image degradations such as noise, blur and scale-variations are known to significantly affect the classification process of textured images. However, due to difficult visual according conditions, such degradation are often prevalent in digital real-world images. We show that these degradations not necessarily strongly affect the discriminative powers of features, in a scenario where similarly degraded images are classified. Contrarily, if the training and the evaluation set contain differently degraded images, the accuracies are decreasing extremely. In this paper, we exploit this knowledge and propose an approach which divides one large database into several smaller ones, each containing similarly degraded images. In order to get sensible database divisions, we use criteria adapted to the respective degradation. In experiments with several degradations, classifiers and feature extraction methods, we show that our method continuously and significantly enhances the classification accuracies.