Research On Heterogeneous Parallel Computing Based On CPU+GPU In Digital Image Processing

With the increasing requirement of speed and accuracy in the man-machineinteractive system, processing time of machine vision is demanded shorter and shorter.Therefore, it’s difficult to meet the speed requirement singly using CPU for computation.Recent years, with the development of GPU hardware, GPU has shown high power offloating-point arithmetic and high bandwidth, it’s widely used for general calculation.However, the developers have to complete the program using the computer graphicapplication interface; it also makes great difficulties for non-computer professionals.NVIDIA proposed CUDA architecture allows developers to program using GPUdirectly without packing the data into texture, and the image processing algorithms canbe ported to GPU using CUDA C language, reducing the difficulties of development.The paper first reviews the development of GPU hardware and describes thegeneral situation of CUDA, programming model, storage model and application. Thenthe compiling process of CUDA program is analyzed. Finally we research on severalimage algorithms whether they are parallelism and propose the corresponding parallelalgorithms based on GPU, and then give a detail comparison of data and results.The main works of this thesis are listed below:(1) The paper analysis the image interpolation algorithm and cubic convolutioninterpolation algorithm, then proposes the realization of two algorithm based on GPU.Finally, the experimental results of all platforms are analyzed.(2) This paper analysis the mask pattern smoothing algorithm in image smoothingand Roberts edge detection algorithm in image segmentation and proposes the realizationof these algorithm based on GPU. Finally, the experimental results are analyzed.(3) Proposed two optimization schemes for bilinear interpolation algorithm andaccelerated the bilinear interpolation procedures adopting texture memory and CUDAasynchronous execution. We compare the unoptimized results with optimized resultsto verify the correctness of the program.