Co-occurrence-based texture analysis using irregular tessellations

Grey level co-occurrence features are one of the most powerful feature sets available for texture analysis. However, the moving window commonly employed to define the statistical scale at which the co-occurrence matrix is obtained assumes spatial stationarity of the underlying random field. This assumption is inappropriate in the case of natural images and may result in the mixing of different structures at various positions that can yield misleading features, affecting any subsequent analysis or classification. To minimise this problem, we present a method for obtaining co-occurrence features from the irregular tessellation of an image. Such tessellation is considered to be the result of a filtering or pre-segmentation step guaranteeing a certain degree of homogeneity within each tessellation element, and thus offering a more optimal statistical scale at each location in the image. Experimental results and a comparison between features obtained from various irregular and square tessellation elements in a set of natural texture images are presented. They show that features obtained with our method have a similar behaviour to those generated from a traditional square window.