Classification of natural textures by means of two-dimensional orthogonal masks

An approach to texture classification in which each local neighborhood is transformed reversibly into a minimal set of uncorrelated features is described. These features are found to represent local average, as well as first- and second-order derivative information of the texture. Comparison to other techniques of similar complexity, using identical classification procedures, shows the superiority of the proposed approach. Moreover, this approach leads directly to a real-time parallel implementation.<>