Research On Heterogeneous Computing-based Image Compression Technologies

Image compression technology is one of the important research in image processing field. Image compression technology is used widely in many areas, such as digital photography, mobile multimedia, medical imaging and so on. The image compression technology can effectively reduce the redundancy information and assure the quality of a digital image. That make image storage and transmission more efficiently. Since the time complexity of compression algorithm and the amount of image data increased, the existing compression CPU-based algorithm not meet the real-time requirements any more. With GPU computing power and programmability improvement continuously, the use of GPU for general computing purpose has become a hot research. Therefore, CPU+GPU heterogeneous computing provide a new solution in image compression coder.At first this paper described the heterogeneous computing technology and elaborated the hardware framework and software model of CUDA technology, including the relationship between host and device, the kernel function, the thread hierarchies and the storage model. Secondly, this paper also summarized the CUDA parallel programming method. Thirdly, the image compression technology was introduced, focusing on the JPEG algorithm, including the principle, function and algorithm of color space conversion, discrete cosine transform, quantization and entropy coding.In order to improve the speed and performance of the image compression algorithm, a heterogeneous computing-based parallel image compression algorithm was proposed. Basing on the original CPU-based optimization algorithms and considering the characteristics of the GPU, a parallel algorithm was came up. In the color space conversion step, using a float multiplication instead of look-up table method could avoids calculating IO bottleneck. In the DCT step, the two-dimensional DCT algorithm was optimized by the symmetry and a heterogeneous computing-based rapid transformation algorithm was realized. In the quantization step, according to the characteristics of JPEG that most of the high frequency components after quantization will be zero, only first 32 components were quantized so that a fast algorithm was achieved. Finally, a reasonable computing amount of CPU and GPU was allotted to avoid the short board. Experimental results show that, in the case of less different of PSNR and compression quality meeting the needs, compared with the serial version the parallel version won above 25 times performance increase. The parallel algorithm has a theoretical and practical value.