FEATURE VALUE SMOOTHING AS AN AID IN TEXTURE ANALYSIS.

When texture features are measured on small subimages, they are unreliable, but if we use large subimages, it is hard to find subimages that are uniformly textured. A compromise approach is described: measure the features on small subimages and smooth the resulting feature values in such a way that neighboring subimages that belong to differently textured regions are unlikely to influence one another. When this is done, classification performance improves substantially. Improvement is also obtained when the subimages are classified probabilistically, and relaxation is used to adjust the class probabilities. The problem of choosing a window size that minimizes overall misclassification probability is also discussed.