Refereed full papers (journals, book chapters, international conferences)

1997

Texture classification is in essential a problem of finding an optimal mapping from texture feature space to texture class space. In general, the a priori knowledge of textures in real problems is described in terms of expert rules and in terms of known samples. Statistical classification systems or classical fuzzy logic systems have difficulty to integrate both kinds of knowledge. In the present paper, we propose neural fuzzy systems (NFS) using asymmetric pi membership functions (APF), their learning algorithm based on a new global optimization criterion and the application to texture classification. The NFS using APF shows the following advantages: (1) The APF gives a more general model of fuzzy rules, which improves the precision of NFS. (2) The smoothness of APF assures a good convergence of the system, which avoids oscillations in learning. (3) Based on the new global optimization criterion, the NFS can integrate both the expert knowledge in terms of fuzzy rules and the numerical training data for system learning, which is difficult for classical multilayer network or FLS. (4) The learning algorithm is simple, which is similar to that of classical multilayer network. (5) The NFS permits a refinement of the initial expert knowledge, and the new fuzzy rules found are easy to interpret. (6) When more training data are available in the future, a new training will demand less time of learning and realize a natural forgetting effect for non-precise initial expert knowledge and ancient data, which is desired for many intelligent systems. The NFS using APF is implemented and applied to texture classification, experimental comparison with other methods shows the good performance of such systems.