Genetic Vector Quantisation for Digital Image Compression

In this paper, we describe a novel approach to image quantisation by using a genetic algorithm to find near-optimal code words. Moving away from the traditional binary string representations of genetic data, the proposed system applies genetic operators directly on pixel blocks thus avoiding any conversion to a conventional format. To operate on the data this way, operators for fitness, combination and mutation have been tailored to produce near optimal results in terms of high visual and statistical quality. To benchmark the system, we chose the popular Competitive Learning Neural Network because in terms of format, data and operation it is the closest stochastic technique to the proposed algorithm. The ability for the Genetic Algorithm to reach a near global optimum is demonstrated by the quality of the reconstructed images produced, when compared to the Neural Network for the same compression ratios. A key feature of the Genetic Algorithm is the ability to produce codebooks that contain sufficient words to fully convey detail, flat regions and shading. On the other hand the Competitive Learning Neural Network, when used under the same settings, produces high levels of detail at the cost of heavy artefacts in smooth and shaded areas. When both compressors are measured by PSNR values of their reconstructed image quality, the proposed algorithm constantly outperforms the benchmark for a group of test samples, which is evidenced by our extensive experimental reports.