Multi-scale discriminant analysis and recognition of signals and images.

摘要

A successful pattern recognition scheme starts with efficient extraction of the most discriminant information elements from various, possibly imprecise, sources, followed by an intelligent combination of this information in a context dependent framework of low complexity.; Conventional multiscale basis selection and feature extraction based on compression and approximation-based criteria are not necessarily the best approaches for classification and segmentation purposes. Instead, a class separability based approach is preferable. In this thesis, we explore methodologies for lower-dimensional adaptive multi-scale discriminant basis selection. Depending on the task, these methodologies are applied to local windows or to the whole pattern. Our tools in this analysis are derived from theories of wavelet packets and multi-scale local bases on the one hand, and from the statistical theory of discriminant cluster analysis on the other hand. The goal is to find efficient multi-scale representations that yield maximum between-class separations and minimum within-class scatters.; We also investigate the effectiveness of soft decisions in representing the vagueness, uncertainty and imprecision of the classification sources. Based on the principle of least commitment in designing pattern recognition and consensus-theoretical concepts, we try to improve the reliability of our classification system through integration of soft decisions obtained from various observations and/or sources. The combination of decisions is based on the discrimination power of each source and the relevance to the current observation. We use ideas from consensus theory, fuzzy neural learning, and evidential reasoning.; Our methods of multi-scale local/global basis selection and context-dependent decision integration are applied to in several different domains, including texture and document image classification and segmentation, radar signature classification, and human face recognition. The results show that superior or highly competitive performance can be obtained using small feature sets and simple classifiers. The resulting systems are typically of low complexity and, since no iterative computations are involved, most of the calculations can be done in parallel. The proposed ideas can be extended in several directions and can be applied to many pattern recognition and segmentation tasks.