Speeding up fractal image compression by working in Karhunen-Loeve transform space

The main weakness of fractal image compression is its long encoding time needed to search the entire domain pool to find the best domain-range mapping. To solve the problem, some solutions were proposed but most of them do not employ neural networks (only the use of Kohonen SOM for clustering was reported). The paper proposes a new method based on Karhunen-Loeve transform (PCA networks), which attempts to use neural networks' well-known adaptability in order to find a good feature vector for a block. Performance regarding network generality, quantization of the transform coefficients, comparison with DCT and kd-tree search, was explored. Results prove that the proposed method slightly outperforms state-of-the-art methods.