On texture classification

Texture analysis has found wide application in, say, remote sensing, medical diagnosis, and quality control. There are many ways to classify image texture and many approaches split the problem into extraction followed by classification. We describe feature extraction using the new Statistical Geometrical Features in comparison with Liu's features, features from the Fourier transform using geometrical regions, the Statistical Grey Level Dependency Matrix and the Statistical Feature Matrix. We also include a formal analysis concerning rotational-invariance in the Statistical Geometric Features. Classification techniques considered here include the K-Nearest Neighbour Rule, the Error Rack-propagation method and the new Generating-Shrinking Algorithm. A particular consideration is scale-invariance in the feature space since this implies that textures can be classified as the same, even when the overall illumination level differs. Experimental evaluation on the whole Brodatz texture set shows that the Statistical Geometrical Features can give the best performance for all the considered classifiers, that the Generating-Shrinking Algorithm can offer better performance over the Err or Back-Propagation method and that the K-Nearest Neighbour Rule's performance is comparable with that of the Generating-Shrinking Algorithm. Also, the combination of the Statistical Geometrical Features with the Generating-Shrinking Algorithm constitutes one of the best texture classification systems considered.