Scalable parallel wavelet transforms for image processing

Algorithms for 2D wavelet transform decomposition on clusters of workstations are described and analyzed. For the parallel algorithm employed, the computation of the transform is structured so that the exchange of intermediate transform coefficients is restricted only to neighboring processors and the amount of data communicated is independent of the problem size. Results show that the performance of the parallel implementation improves with increasing data size making the parallel algorithm particularly suitable for applications such as image processing, image coding and computer vision. Timings measured on a Myrinet connected Beowulf cluster agree well with the theoretical analysis and indicate that the implementation is cost optimal.