Research On Parallel Method For Image Denoising Via Sparse Representations

Digital images are often degraded by particular image noise during image capturing, storing and transferring; those image noise resultes in image distortion, so image denoising is a critical problem in image process. The purpose of image denoising is to remove image noise as much as possible for improving image quality, preparing for image recognition and image understanding. In recent years, sparse representation theory is paid much attention and has been successfully applied to image denoising. CUDA (Compute Unified Device Architecture) is a parallel computing architecture developed by NVIDIA Company. This architecture can largely improve computational performance by using processing capability of GPU.In traditional sparse decomposition denoising algorithms, denoising effects were merely taken into account, but computational efficiency of these traditional algorithms was often neglected. So efficiency of these traditional algorithms was often low. In the CUDA, CPU and GPU can work collaboratively in parallel computing, as a result, the efficiency of algorithm can be significantly increased. An improved parallel image denoising algorithm based on CUDA via sparse decomposition is proposed in this thesis, the experiments results showes that the improved algorithm can largely improve image denoising algorithm's computational performance.In this thesis, three kinds of sparse decomposition algorithm respectively based on DCT, training and adaptive dictionary were analysed. In the procedure of sparse coefficients solution in sparse decomposition algorithm, those time-consuming parts are specially analysed and put in feasible analysed concern parallel computing, as a result a concurrent strategy is proposed. In the concurrent strategy, CUDA was mainly used in some independent modalities in algorithm, this make independent tasks simultaneously perform. This algorithm was encapsulated and enforced utility software.Major work including:(Ⅰ) Principle and implementation sparse decomposition algorithm are introduced, and image denoising procedure based on DCT, training and adaptive dictionary were analysed in detailed, and those can put in parallel computing are pointed out.(Ⅱ) A new improved parallel method based on DCT, training and adaptive dictionary algorithms was proposed and implemented by CUDA; the experiment results showes that this improved method has a higher efficiency.(Ⅲ) The parallel improvement algorithm was encapsulated by CUDA, and implemented relevant software, which were useful for the method's appliance.In this thesis, for improving computational efficiency of sparse decomposition image denoising algorithm, parallel computing mechanism of CUDA supplies a better approach, instance and benefit reference for parallel improvement of other analogy algorithms.