A standard technique for improving the performance of data compression algorithms is to pipeline them. For instance, the ubiquitous JPEG standard, used for image compression, is more or less a pipeline of stages performing operations such as discrete cosine transform (DCT), quantization, or entropy encoding. The authors suggest using this philosophy for a wavelet-based compression algorithm similar to JPEG2000, namely wavelet difference reduction (WDR), by applying a prior boosting phase based on singular value decomposition (SVD) on images, seen as high-dimension matrices.

Even though SVD, in itself, is a lossless transform, further compression can be achieved by selecting only a significant subset of the largest non-zero coefficients of the diagonal matrix part of the SVD transform of an image. The inverse SVD transform of these pruned matrices is then performed to yield a new image, to which WDR is applied. The authors show that this two-stage technique is quantitatively superior to WDR and JPEG2000 by about 5 dB (peak signal-to-noise ratio) or ten percentage points (structural similarity), depending upon the compression ratio.

This rather short and easy-to-read paper introduces a new trick to obtain incremental improvements to existing compression techniques, particularly wavelet-based ones. Although not earth-shattering in itself, it should nonetheless be of interest to researchers specialized in improving image processing algorithms.