Research Of Digital Image Processing Parallel Algorithms Based On GPU

Digital Image Processing is the core part in many different areas. With theincreasing of the image size and the pixels dignity, the demanding of the time costincreases at the same time. The traditionalalgorithms based on CPU can’t satisfy thereal-time or near real-time processing speed. Therefore the processing speed of digitalimage is facing enormous challenges.Traditionally the application of GPU is limited in graphics rendering computingtasks. This makes great extravagance of the computing resource. With thedevelopment of GPU,the research of GPGPU has become more and more.TheGPGPU is based on CPU+GPU hybrid execution model, which the GPU isresponsible to deal with the problems that can be expressed as data-parallelcomputations while the CPU is responsible to deal with those are complex and notsuitable to be expressed as data-parallel computations. GPU provides a low-price andefficient platform for digital image processing.At first,the paper introduced the development process of GPU、compare betweenGPU and CPU、structure of GPU. Then, the paper mainly introduce the platform ofCUDA, including the CUDA programming model、the requirements for hardware andsoftware environments and memory structure. Furthermore, the paper represents theoptimized design of three digital image processing algorithms including the digitalimage filtering、image motion measurement based on optical joint transformcorrelator、star image registration. We proposed a parallel model of GPU for thetraditional algorithm and used Visual Studio2010、OpenCV2.3.1、CUDAprogramming language to realize it. Experiment result shows that the parallel modelhas a speedup compared with the serial CPU program.At last the paper introduces the method to invoking the CUDA kernel by usingMatlab. Defining a CUDA kernel in a mex function in Matlab can accelerate thespeed of digital image processing more easily as the Matlab is much easier than C language.