This dissertation investigates computational issues of feature extraction and image organization at different levels. Boundary detection and segmentation are studied extensively for range, intensity, and texture images. We developed a range image segmentation system using a LEGION network based on a similarity measure consisting of estimated surface properties. We propose a nonlinear smoothing algorithm through local coupling structures, which exhibits distinctive temporal properties such as quick convergence.
We propose spectral histograms, consisting of marginal distributions of a chosen bank of filters, as a generic feature vector based on that early steps of human visual processing can be modeled using local spatial/frequency representations. Spectral histograms are studied extensively in texture modeling, classification, and segmentation. Experiments in texture synthesis and classification demonstrate that spectral histograms provide a sufficient and unified feature in capturing perceptual appearance of textures. Spectral histograms improve significantly the classification performance for challenging texture images. We also propose a model for texture discrimination based on spectral histograms which matches existing psychophysical data. A new energy functional for image segmentation is proposed. With given regional features, an iterative and deterministic algorithm for segmentation is derived. Satisfactory results are obtained for natural texture images using spectral histograms. We also developed a novel algorithm which automatically identifies homogeneous texture features from input images. By incorporating texture structures, we achieve accurate texture boundary localization through a new distance measure. With extensive experiments, we demonstrate that spectral histograms provide a generic feature which can be used effectively to solve fundamental vision problems.
Based on a novel and biologically plausible boundary-pair representation, perceptual organization is studied. A network is developed which can simulate many perceptual phenomena through temporal dynamics. Boundary-pair representation provides a unified explanation of edge- and surface-based representations.
A prototype system for automated feature extraction from remote sensing images is developed. By combining the advantages of the learning-by-example method and a locally coupled network, a generic feature extraction system is feasible. The system is tested by extracting hydrographic features from large images of natural scenes.
Title and Table of Contents [PDF format (123 KB)] [gzipped PS format (94 KB)]
Chapter 1. Introduction [PDF format (116 KB)] [gzipped PS format (110 KB)]
Chapter 2. Range Image Segmentation Using a Relaxation Oscillator Network [PDF format (2.1 MB)] [gzipped PS format (637 KB)]
Chapter 3. Boundary Detection by Contextual Nonlinear Smoothing [PDF format (633 KB)] [gzipped PS format (1.2 MB)]
Chapter 4. Spectral Histograms: A Generic Feature for Images [PDF format (1.6 MB)] [gzipped PS format (3.6 MB)]
Chapter 5. Image Segmentation Using Spectral Histograms [PDF format (710 KB)] [gzipped PS format (2.0 MB)]
Chapter 6. Perceptual Organization Based on Temporal Dynamics [PDF format (388 KB)] [gzipped PS format (172 KB)]
Chapter 7. Extraction of Hydrographic Regions from Remote Sensing Images
Using an Oscillator Network with Weight Adaptation
[PDF format (1.4 MB)] [gzipped PS format (4.8 MB)]
Chapter 8. Conclusions and Future Work [PDF format (191 KB)] [gzipped PS format (840 KB)]
References [PDF format (65 KB)] [gzipped PS format (62 KB)]