Optimization of a texture measure for image classification

Automated image classification is often aided by incorporating a texture measure to distinguish patterns of variation in image spectral values. Several such measures are in common use, with each being based on a different abstract theory of texture. In employing one of these measures, it is never certain that the given texture theory is well-suited to the particular images at hand. To investigate this, a technique was developed for optimizing a texture measure for the best classification of a training image. The measure was adapted from the texture energy theory developed by K. I. Laws in 1980. The optimizations were carried out using the method of simulated annealing to search for an optimum point in the space of matrix filter masks used in computing texture energy. Texture measures were optimized for the training image, and classification performance was then compared with Laws' measures for the training image and two other images from the same geographic region. The results showed that randomly initialized texture measures could be optimized to give performance matching the best obtainable with texture energy. In fairness, it must be noted that none of the optimizations reached levels far in advance of this performance. This in itself may be instructive, demonstrating the robustness of the texture energy measure, and the diminishing returns in pushing beyond its level of performance.