查看文章

In this paper we address the problem of color texture classi cation and present results on two practical problems. The central idea is to combine color and texture information through the multiresolution decomposition of each channel in order to take as classi cation vector the energies and cross correlations of the coe cient images. However, this simple approach can be materialized in many di erent ways, as a several decisions have to be taken, each one allowing multiple choices: the multiresolution decomposition scheme (for instance, Mallat's, a trous, wavelet packets), the subspaces base family (and within it, which speci c base), number of decomposition levels, space for color representation and nally, the classi cation features to be computed from the decomposition. Instead of simply trying some possibilities and take the best one, we have assessed a very large number of combinations, trying to nd out which are the important and the non {relevant issues with regard the classi er performance. In addition, we propose three image models as a framework for color texture classi cation, depending on how texture is combined with color. This allows us not only to initially select the appropriate types of features but also to reduce the number of classi cation parameters so that the training set does not need to be large. This framework has been successfully applied to two speci c machine vision problems, namely, the sorting of ceramic tiles into perceptually homogeneous classes and the recognition of metalized paints for car re nishing.