MDL-based band selection and adaptive penalties for hyperspectral image segmentation

Image segmentation is the process of assigning each pixel in an image to a class, such that all pixels of a given class are similar, but the statistics of each class differ. Markov random fields (MRF) are widely used as a spatial prior, so the segmentation will have a regular spatial structure. This dissertation presents feature-extraction and selection algorithms and an adaptive penalty parameter-estimation procedure to be used in MRF segmentation.; Hyperspectral images are multispectral images with such a large number of bands that each pixel's data vector approximates its continuous spectrum. These images are remarkably rich in information, but they also have severe redundancies and are extremely large. Therefore, we combine feature extraction and selection with image segmentation in the following five-step procedure for hyperspectral images: (1) a global wavelet feature-extraction and selection algorithm inexpensively removes many redundant bands, (2) principal components is used on 64 x 80-pixel blocks to refine feature extraction and selection locally, (3) blockwise MRF segmentation with cluster splitting finds the local class set, (4) merge the blocks and their classes, (5) resegment to remove block artifacts. The algorithm generally performs well on real and synthetic data, except that Step 4 is needlessly complicated and makes some improper merges. Several methods for correcting this are suggested.; All MRF image-segmentation criteria as in Step 3 have spatial-penalty parameters that must be chosen. An adaptive algorithm that chooses the penalty parameters to maximize the pseudo-likelihood (PL) of the current image was developed by Lakshmanan and Derin, but it uses a costly simulated-annealing algorithm. We use a decoupling argument to find simple, closed-form solutions for the PL penalty parameters of a globally adaptive (GA) MRF criterion with boundary and region penalties. A theoretical analysis shows that GA penalties only minimize the error rate if the scene has certain weak symmetry properties. For example, all boundaries must be equally rough. This is not always satisfied in practice, so we also introduce an MRF with class-pair-conditional (CP) boundary penalties. We segment both synthetic and real images to validate the theoretical analysis and illustrate the capabilities and limitations inherent to the PL approximation.