Image Compression by Approximated 2D Karhunen Loeve Transform
Image compression is performed by 8 x 8 block transform based on approximated 2D Karhunen Loeve Transform. The transform matrix W is produced by eight pass, modified Oja-RLS neural algorithm which uses the learning vectors creating the image domain subdivision into 8 x 1 blocks. In transform domain, the stages of quantisation and entropy coding follow exactly JPEG standard principles. It appears that for images of natural scenes, the new scheme outperforms significantly JPEG standard: at the same bitrates it gives up to two decibels increase of PSNR measure while at the same image quality it gives up to 50% lower bitrates. Despite the time complexity of the proposed scheme is higher than JPEG time complexity, it is practical method for handling still images, as C++ implementation on PC platform, can encode and decode for instance LENA image in less than two seconds.
KeywordsDiscrete Cosine Transform Image Compression Quantisation Scale Entropy Code Quantisation Matrix
Unable to display preview. Download preview PDF.
- 2.Cichocki A., Kasprzak W., Skarbek W.: Adaptive learning algorithm for Principal Component Analysis with partial data. Proceedings of 13-th European Meeting on Cybernetics and Systems Research. Austrian Society for Cybernetics Studies, Vienna, Austria (1996) 1014–1019Google Scholar
- 3.Diamantaras K. I., Kung S. Y.: Principal component neural networks — Theory and applications. John Wiley & Sons, Inc. (1995)Google Scholar
- 4.Jain A.K.: Fundamentals of digital image processing. Prentice-Hall International, Englewood Cliffs NJ (1994)Google Scholar
- 5.Mitchell, Pennebaker: The JPEG standard (1995)Google Scholar
- 8.Skarbek, W., Sikora, R., Pietrowcew, A.: Modified Oja-RLS Algorithm — Stochastic Convergence Analysis and Application for Image Compression. Fundamenta Informaticae. (to appear in 1999)Google Scholar