English

Abstract

Texture classification is one of the most studied and challenging problems in computer vision. A key re- quirement of successful texture classification algorithms is their ability to quantify the complex nature and diversity of real world textures. Recent developments in automatic texture classification have demonstrated the effectiveness of modeling texture elements as cluster centers of responses of a filter bank. Such methods rely primarily on similarity measurements of frequency histograms of vector quantized versions of the target texture. A main problem with these approaches is that pure frequency histograms fail to explicitly account for important spatial interaction between learned texture elements. Spatial interaction is key to classification when analyzing textures that have similar texture element frequency but differ in the way the texture elements are distributed across the image. In this paper, we propose the use of co-occurrence statistics to account for the spatial interaction among learned texture elements. This is accomplished by calculating spatial co-occurrence statistics on the maps of the learned texture elements. We demonstrate the effectiveness of our method on images from the Brodatz album as well as natural textures from a tropical pollen database. We also present a comparison with a state-of-the-art method for texture classication. Finally, our experiments show that the use of spatial statistics help improve the classification rates for certain textures that present sparse and statistically non-stationary texture elements such as pollen grain textures.