Hierarchical maximum entropy partitioning in texture image analysis

This paper presents an effective texture representation which captures the statistics (or distributions) of structural and/or spatial relations of grey levels within certain neighborhood in a texture image. The structural and/or spatial relations are captured by various feature extraction operators to generate feature images. Then, the joint distributions of the features which we termed feature frequency matrices (FFM) provide the statistics and representation of the texture image. A partitioning scheme to 'compress' the FFM such that only relevant information is retained is proposed. The partitioning scheme is based on the hierarchical maximum entropy discretization scheme which minimizes the loss of information. The efficacy of the representation is demonstrated using homogeneous texture images.