A Scalable Implementation Of Adaptive Directional Lifting Wavelet Based Image Decomposition On CUDA

Image compression methods based on wavelet transformation has always been an important research direction in image processing field. And lifting scheme can easily construct wavelet mapping from integer to integer. The adaptive directional lifting wavelet (ADL) in which the texture direction of the image can be detected on the fractional pixel precision, and lifting wavelet transformation can be applied along the direction of the texture. On the other hand, image compression technology based only on single computing core is very limited in terms of efficiency improvement, while parallel computing is an effective way to improve processing speed. CUDA has opened up new avenues looking for high-performance parallel image compression algorithm.Since the ADL algorithm requires interpolation in the transformation process, and also to select the optimal transformation direction, therefore it leads to a larger calculation amount than the general wavelet transform. For the computation-intensive problem in this thesis, a series of researches and improvements have been done, such as analyzing of the execution flow ADL algorithm and evaluating the performance of memory access and computing involved, and then made an comparison after applying some improvements, looking into the inherent parallelism of ADL algorithm etc. And then combined with the characteristics of CUDA such as thread blocks and threads parallelism etc, the defects of existing algorithms were analyzed and corresponding improvements were applied, including separable kernels to avoid unnecessary waiting time between different threads, proper divisions of the computation to eliminating data dependencies between lifting steps, and applying Slice configuration scheme to exploit the parallelism in terms of threads to a larger extend etc. Finally the effect of the original algorithm and the improved algorithm are verified through a series of experiments, and the data obtained from the experiments show that the proposed algorithms have achieved a better speedup, compared to the original algorithm.