CUDA-based Three-dimensional Non-local Mean Filtering Parallel Algorithm Design

Image denoising is an important study area in three-dimension ultrasound imagingsystem. Non-local mean filter is a nice method to remove the speckle noise, and smooththe consecutive area in ultrasound images. However, the algorithm’ computationalcomplexity makes it a highly time-consuming method, and unable to meet therequirement of real-time. Real-time3D ultrasound imaging system requires each moduleof system for data processing in real-time, especially in surgical navigation.Non-local mean filtering method in ultrasound image denoising has a lot ofsuccessful applications, and this thesis has implemented a NLM method based C++. Fordesigning parallel algorithm, how to choose appropriate technology according to thecharacter of both the algorithm and device, and extend from2D to3D images is adifficult problem. In recent years, GPGPU technology shows its advantages in dealingwith large-scale data processing. In2007, Nvidia launched the CUDA (Compute UnifiedDevice Architecture) brought GPGPU into the broader field. Now, CUDA technology hasbeen employed in oil exploration, astronomy, computing, image processing and otherfields. This paper divides the Non Local Mean into several tasks, to confirmthe serial execution and parallel execution part, employs the CPU+GPU heterogeneousmodel. CPU takes charge of the implement of complex logic processing and transactionmanagement, serial parts, and GPU takes charge of compute-intensive large-scale dataparallel computing. In the experiments, the performance of smoothing based2Ddenoising and3D denoising technologies are compared and analyzed, and designed amethod uses CUDA to accelerated the parallel computing method, which efficiency is farhigher than the serial computing, but the specific algorithm implementation must matchwith the hardware platform and design tools. The key of CUDA parallel program designis how to allocate thread-task, common threads, synchronization, and optimizationalgorithms and so on. This paper implements a NLM method accelerated by GPU forreal-time denoising of3D ultrasound images, furthermore, it summarized the usefulexperience for application of CUDA.