GPU Acceleration Of SPIHT Image Compression Algorithm

Image compression is one of the most important researching directions in thefield of information processing, and has a wide range of applications in the imagestorage and image transmission and so on.At present there are many imagecompression methods, mainly comprising compression method based on predictionand transform-based compression method. In the methods of transform-basedcompression, SPIHT image compression algorithm based on wavelet transform is acommonly used image compression algorithm, which has good progressivetransmission characteristics, relatively high peak signal to noise ratio, and supportbothlosscompression and lossless compression, andat the same time has theadvantages of lower computational complexity, easiness of controlling the rate andso on.However, the traditional SPIHT image compression used in the encodingprocess needs the ordered lists, the traversal process of which will take up a lot ofexecution time of the algorithm, resulting in the reduced efficiency of the algorithm,andthus making it difficult to meet some situations that require for the high speed ofprocessing.In order to meet some occasions of efficient image compression, this paperattempts to use CUDA, the GPUhigh performance computing architecture, toachieve SPIHT image compression algorithm.Firstly, after studying the theory ofparallel computing and the CUDA parallel computing architecture,having thein-depth research of the SPIHT image compression algorithm, and analyzing theinternal data dependence relationship of thealgorithm, we propose a parallel GPUimplementation for SPIHT image compression algorithm. The parallel algorithmcontains thewavelet transform algorithm based on block and the parallel SPIHTcoding algorithm based on parallel coding trees, and thus implementsthe paralleldata compression. At the same time, we have also analyzed the data dependencybetween tasks during the process of the SPIHT image compression coding, whichachieves the tasks processed in parallel by making use of the relative independencebetween the different thresholdscoding tasks.Secondly, this paper proposes anapplicable method of transplanting the parallel SPIHT image compression algorithmto theCUDA, andfinished the parallel program coding by using the GPU parallel computing capabilities and the data streams to achieve the task parallelcomputing.Finally, the parallel program was optimized considering the rich memoryresources and I/O transfers of GPU thatmakes the SPIHT image compressionalgorithm eventually toachieve more than8times speedup, thus greatly improvingthe efficiency of the compression algorithm.