A spatial pyramid approach for texture classification

Texture classification, texture synthesis, or similar tasks are an active topic in computer vision and pattern recognition. This paper aims to present two spatial pyramid representations for texture classification. Most techniques designed for texture classification are based on machine learning. Images are usually represented as feature vectors, which are then used to train a classifier. In the spatial pyramid representation, images are divided into increasingly fine sub-regions (bins) and features are extracted from each bin. This representation is able to capture details about the fractal structure of the texture images. Two experiments are conducted on popular texture classification data sets, namely Brodatz and UIUCTex. In the experiments, several kernel representations and kernel classifiers are combined and evaluated. It seems that the spatial pyramid in combination with intersection kernel and Kernel Discriminant Analysis gives the best results. The proposed pyramid representations can improve the accuracy by as much as 5% over the standard feature representation, showing that the pyramid structure is indeed useful for texture classification.