The Study On GPU-based Image Clearness Technology

Image capture and transmission systems are easily interfered by many factors,such as noise from the image sensor, turbulence in the atmosphere, motion blur andunevenness of exposure, resulting in degradation of images. In addition to theimprovements of photoelectric sensors of the front imaging systems, the technologyof image clearness is also necessary for back-end processing. With the increase ofrequirements for precision, performance and speed of clearness, it is urgent to seekquicker and superior algorithms than the traditional methods. Besides, hardwareplatforms of high performance are needed to develop rapidly. Fortunately, theappearance of GPU(Graphics Processing Unit), a new kind of platform which haslower-power and lower-cost advantages than traditional DSP and FPGA platforms,makes it possible. With the launch of GPU-based general purpose computing platform,it is easier for professionals and developers to carry out the research ofhigh-performance parallel computing.Firstly, in this paper, apart from the traditional enhancing and denoising methods,new modeling means, including retinex and dark channel priority algorithm aresimulated and analyzed. The experimental results show that, most of the traditionalalgorithms are computationally intensive; local histogram equalization algorithm maybring block-like effect; homomorphic filtering algorithm is not well adaptive; NLMalgorithm and wavelet algorithm denoise well and are of high parallelism; retinexalgorithm and dark channel priority algorithm are proved to be better in defoggingand color fidelity, except for multi-DOF (Depth of field) and thick-fog images.Secondly, restoration algorithms which are more meaningful than enhancingmethods, such as IBD, NAS-RIF, SA, Richardson-Lucy and class-G fast algorithms,are also tested and analyzed, mainly in stability, robustness and run-time. The resultsindicate that, traditional iterative blind restoration algorithms are of high-complexityand poor-stability. However, the APEX algorithm and mid-frequency blind restorationalgorithm are demonstrated to be of greater restoration-performance, lesstime-consumption and higher parallelism than the traditional ones.Thirdly, an improved NLM parallel algorithm and an improved class-Gmid-frequency blind restoration parallel algorithm are proposed, after analyzing theheterogeneous pattern of CPU-GPU and combing the architecture, storage models and optimization methods of CUDA. Afterwards, both of the algorithms are implementedon the GPU-based platform and gain obvious processing advantages especially forhigh volume of image data. Then, a heterogeneous platform for image clearness is setup, controlling image acquisition card and acquisition channels. After co-development,the system shows higher speed in processing videos.