An image compression algorithm based on the Karhunen Loève transform

The Karhunen-Loève Transform (KLT) is the optimal transform for a block of signal in terms of decorrelation and energy compaction performances. However, it is not used in practice because we need to transmit the transform kernel for every block, which requires huge amount of side information. Hence fixed-kernel discrete cosine transform (DCT) is used instead in many of current image compression algorithms. In this paper, considering a recent trend of requiring high resolution, high quality, multiple view images, we propose an image adaptive transform (IAT) approach where the transform kernels are derived from the KLT for a group of image blocks. The contribution of this paper is to develop a simple method to group the similar image blocks according to DCT coefficient statistics, and also to reduce the number of transform kernels by using the relationship between the DCT coefficients and the corresponding block properties such as edge directions. Since this method needs some amount of side information for transmitting IAT kernels, it does not perform well for the low resolution images at low bit rates, but it outperforms JPEG for the high resolution, high quality and multiple view images that can share the same transforms.