| This dissertation presents a methodology for two-dimensional multi-scale decomposition of textural images. An image is decomposed into a set of component images and a detail image. Each component image consists of primitives of the same size and shape. The combination of all component images forms an approximation image. The detail image is the difference between the original and the approximation images. The algorithm is formulated as an optimization problem that minimizes the number of primitives used under the constraints of completeness, orthogonality, and mean square error. A computer algorithm was developed to implement the decomposition in a computationally efficient manner. The area and count of primitives are shown to be useful texture features. A dominant texture scale derived from the decomposition provides a good reference parameter for computing pixel value co-occurrence features, and run-length features of the approximation image effectively reflect the essence of the underlying texture. The advantages of features based on primitive co-occurrence are demonstrated with real textural image classification. The methodology was applied in extracting texture features of beef muscle images and classifying beef samples into tender and tough categories. The application further shows the usefulness of the decomposition technique. |
