Evolutionary Methodology for Optimization of Image Transforms Subject to Quantization Noise

Lossy image compression algorithms sacrifice perfect imagereconstruction in favor of decreased storage requirements. Modelossy compression schemes, such as JPEG2000, rely upon the discrete wavelet transform (DWT) to achieve high levels of compression while minimizing the loss of information for image reconstruction. Some compression applications require higher levels of compression than those achieved through application of the DWT and entropy coding. In such lossy systems, quantization provides high compression rates at the cost of increased distortion. Unfortunately, as the amount of quantization increases, the performance of the DWT for accurate image reconstruction deteriorates. Previous research demonstrates that a genetic algorithm can improve image reconstruction in the presence of quantization error by replacing the wavelet filter coefficients with a set of evolved coefficients. This dissertation develops a methodology for the evolution of digital filters capable of outperforming the DWT for image reconstruction at a given compression rate in the presence of quantization error. This dissertation compares potential fitness measures for evaluating reconstruction error. Experiments compare the usefulness of local versus standard population initialization and mutation operators. In order to perform an efficient yet thorough traversal of the search space, several recombination operators developed specifically for real-valued evolution are evaluated. Additionally, this dissertation presents and develops a novel technique to emphasize the reconstruction of the high-spacial frequency areas of an image through use of edge detection algorithms and focused evolution. An analysis of the ease of traversal through the fitness landscapes defined by various image quality measures supports the development of a framework for evolving robust image transform filters. Particular emphasis is placed upon the development of transforms that provide consistently accurate reconstruction of quantized satellite and aerial reconnaissance images. The development of transforms that preserve the intelligence that can be gained from highly compressed images transmitted over a limited bandwidth is of defense and security interest. This dissertation assembles a database of publicly available satellite images collected for a wide range of subjects, including aircraft and airfields, naval bases, army bases, cities, and factories. Experiments employing these images are geared toward the development of filters appropriate for civilian and military aerial reconnaissance applications requiring limited bandwidth for image transmission. Because the evolution employs the DWT algorithm, the resulting filters are easy to implement in hardware appropriate for digital signal processing applications