The Research Of Parallel Non-local Means Denoising Algorithm

An image will be inevitably affected by noise in the process of acquisition and transmission, the noise will drop down the image quality, so in order to effectively obtain the useful information of the image, image denoising is particularly important. There have been many denoising methods, in which, the non-local means(NLM) denoising algorithm is widely used because of its simplicity and outstanding denoising performance, however, the traditional serial method is too time-consuming when processing a large amount of data of image, in order to reduce the denoising time of the NLM algorithm, this thesis mainly study how to parallel the NLM denoising algorithm. In this thesis, our main research work is as follows:(1) We implement a single GPU parallel NLM denoising algorithm based on Open CL parallel framework, in this method, we put all the pixels of the image on each processing unit of GPU, use Open CL parallel framwork to caculate the denoising result of each pixel on GPU, it can realize a higher speedup than serial NLM denoising algorithm, then we verify the denoising effect of the parallel algorithm by the Peak Signal to Noise Ratio, the experimental results show that the denoising effect of serial NLM algorithm is as same as the denoising effect of parallel NLM algorithm. Then we verify the performance portabolity of the single GPU parallel NLM denoising algorithm based on Open CL on different AMD CPU platforms and different GPU platforms such as NVIDIA FX4800 and NVIDIA Tesla k20. In order to furtherly improve the efficiency, on the basis of single GPU parallel NLM denoising algorithm based on Open CL, we put forward and design a faster parallel algorithm based on double GPU under Open CL framewok, the experimental results show that, when the image size increase, the speedup of the parallel algorithm based on double GPU under Open CL framewok is easier to reach two times than the speedup of the parallel algorithm based on single GPU under Open CL. Finally, we use Qt software to encapsulate the serial algorithm, single GPU parallel algorithm and double GPU parallel algorithm together and make it a software system.(2) We proposed a parallel NLM denoising algorithm based on Open MP, the experimental results on 24-cores Xeon CPU show that, to image size of 16384×16384, the parallel NLM denoising algorithm based on Open MP can get a speedup of more than 20 times than serial NLM denoising algorithm. Then we do some experiments on MIC, we compare and analyze the acceleration effect of offload mode and native mode on MIC, and we respectively compare and analyze the acceleration performance of parallel NLM denoising algorithm based on Open CL and parallel NLM denoising algorithm based on Open MP on multi-CPU and MIC.