1 Non - Parametric Texture Learning

Texture is one of the most informative visual cues that help us understand our environment. Texture analysis is an important step in many visual tasks, such as scene segmentation, object recognition, and shape and depth perception. In this chapter we consider the problem of texture recognition and provide an overview of our recent work on this topic ((21, 19, 18]). Our method is based on representing textures in frequency and orientation space, and using non-parametric learning schemes for clas-siication. We present state-of-the-art recognition results on a 30 texture database and compare the performance of a rule-based network, the k-nearest neighbor and feedforward neural-network classiiers. An important extension to the system allows for rotation invariance. Experimental results are presented for large-database rotation-invariant natural texture recognition. 1. Introduction Textures in an image are usually very apparent to a human observer (for example see Fig. 1), but no simple mathematical deenition captures all aspects of the very diverse texture family. It is this lack of good models that makes automatic description and recognition of these patterns a complex and, as yet, an unsolved problem. Although researchers approach texture diierently, most would agree that the texture family can be categorized into two main categories | structured (or oriented) and unstructured (non-oriented) textures. We can loosely identify textures as structured when there are clearly deened microstructures with a speciic repetitive ordering. The unstructured textures are more fractal-like or stochastic in their appearance. Much eeort has been expended to automatically segment and recognize the diierent types of natural textures. We will brieey mention next some of the classical methods in the literature.