Multi-Image Classification And Compression Using Vector Quantization

Abstract

Vector Quantization (VQ) is an image processing technique based on statistical clustering, and designed originally for image compression. In this dissertation, several methods for multi-image classification and compression based on a VQ design are presented. It is demonstrated that VQ can perform joint multi-image classification and compression by associating a class identifier with each multi-spectral signature codevector. We extend the Weighted Bayes Risk VQ (WBRVQ) method, previously used for single-component images, that explicitly incorporates a Bayes risk component into the distortion measure used in the VQ quantizer design and thereby permits a flexible trade-off between classification and compression priorities. In the specific case of multi-spectral images, we investigate the application of the Multi-scale Retinex algorithm as a preprocessing stage, before classification and compression, that performs dynamic range compression, reduces the dependence on lighting conditions, and generally enhances apparent spatial resolution. The goals of this research are four-fold: (1) to study the interrelationship between statistical clustering, classification and compression in a multi-image VQ context; (2) to study mixed-pixel classification and combined classification and compression for simulated and actual, multispectral and hyperspectral multi-images; (3) to study the effects of multi-image enhancement on class spectral signatures; and (4) to study the preservation of scientific data integrity as a function of compression. In this research, a key issue is not just the subjective quality of the resulting images after classification and compression but also the effect of multi-image dimensionality on the complexity of the optimal coder design.