Study On GPU-based High Performance Decoding Methods For Remote Sensing Images

The large data volumes, the strict requirement of real time response, make it of great importance to perform real time, efficient compression for the remote sensing images on the satellite. Currently, it is conventional to use space borne embedded hardware processing platforms assembled with space level, high reliable components for the compression of data, which is determined both by the small room, low power supply of the satellites and by the severe radiance environment of the space. In comparison with the encoding environment, the situation for the ground decoding system is much better, which has brought possibilities for various solutions. Traditional solutions use super computer clusters for the decoding system to ensure the performance. However, the cost of maintaining such super computers is prohibitively high, as separate machine rooms with proper temperature are required, and the power consumption is huge. Nevertheless, the newly appeared Graphics Processing Unit(GPU) greatly solved the contradiction between computational ability and cost, becoming an attractive solution for high performance application systems.Thus, the GPU-based high performance computing solutions are selected to handle the problems encounterred in the decoding of images in remote sensing. In the study, two widely used algorithms, the Set Partitioning in Hierarchical Trees(SPIHT) and the JPEG2000 are concentrated on, as the SPIHT-based hardware encoding system developped by our team has received successful applications in several satellites launched by our country, and the JPEG2000-based high performance encoding chip designed by our team will also be used in future remote sensing missions. The acceleration methods of Discrete Wavelet Transform(DWT) are researched first as the DWT is adopted by both algorithms when performing the time-frequency analysis, and the computational complexity of the DWT usually makes it the bottleneck in the decoding systems. Then, the GPU-based fast DWT is applied to system level decoding systems, the SPIHT decoding system for the Chang E-II satellite and the JPEG2000 decoding system for the high performance encoding chip. In the SPIHT decoding system, an 83 x speedup is achieved by further combining the computational ability of the multi-core CPUs using a heterogeneous parallel model; in the JPEG2000 decoding system, the cooperation model between CPU and GPU is further studied and a speed of 166 fps is achieved. The main work in this thesis is summarized as follows:(1) A GPU-accelerated DWT method based on row-column transposition and a method with fast column transform are proposed. In the first method, the massive parallel execution elements of the GPU are used to accelerate the DWT to solve the speed problem; in the second method, the coalesced globle memory accesss is used to improve the memory access efficiency.(2) A GPU-accelerated DWT method using the block-based approach is proposed. The methods proposed in(1) are vulnerable when the image size is too large as too much shared memory is required. Therefore, the images are partitioned into overlapped small blocks and the DWT is performed on each block in the proposed block-based method. Experimental results have verified that the block-based method reduced both the amount of shared memory usage and the access of global memory, thus a better speed is achieved.(3) A CPU-GPU heterogeneous parallel computing architecture is proposed for the SPIHT decoding system. The existing serial decoding method used in the SPIHT decoding system cannot meet the real time requirements in some special applications. On the basis of fast GPU-based DWT, the parallel computational ability of multi-core CPUs is also used to boost the decoding speed with a pipeline structure. The system finally achieved an 83 x speedup and a speed of 91 fps, which was successfully used in the Chang E-II decoding system.(4) An inter-frame parallel decoding method for JPEG2000 using CPU-GPU collaborative operation is proposed. The pipeline structure requires that the speeds of the CPU part and the GPU part are matched and the complexity is high. Therefore, an even general cooperating model between the CPU and the GPU is proposed. Moreover, the inter-frame parallel scheme is chosen as existing intra-frame parallel methods have the disadvantages of low parallel degree, low efficiency and low expansibility. In the CPU-GPU collaborative inter-frame parallel decoding system, a 20 x speedup, which means, a speed of 166 fps, was achieved. The system was also successfully applied to the high performance decoding system for the JPEG2000 encoding chip.