Evolved Transforms for Improved Image Compression and Reconstruction under Quantization

Previously reported research efforts demonstrated that a genetic algorithm can evolve coefficients describing transforms that outperform standard wavelets, by reducing the mean squared error (MSE) apparent in reconstructed signals under conditions subject to quantization. This paper describes new results that substantially improve the state-of-the-art in evolved transform performance. Matched forward and inverse transform pairs trained against selected images consistently reduce MSE by more than 22% (1.126 dB) when applied to an arbitrary population of similarly quantized test images, yet still achieve the same amount of compression. Key-Words: wavelets, genetic algorithms, image compressions, quantization