The Research Of Video Parallel Dehazing Algorithm Based On CUDA

The method of single frame image defogging based on dark channel prior,because of its solid mathematical mechanism,can obtain a clear and natural image.However,in the practical application,the algorithm required to use the soft matting to refine the raw transmission map in order to suppress the halo effect.But the calculation of soft matting algorithm was complex,high time-consuming,therefore,improving the speed of dark channel prior algorithm was a key problem.This paper analyzed the reason of foggy image degradation and fuzzy mechanism,and proposed a fast dark channel prior defogging algorithm based on Median-modified Wiener filter(MMWF),using GPU to implement parallelization of the method on the platform of CUDA,and the parallel design of the video defogging algorithm on the GPU side was carried out.Specific research works of this paper can be summarized as follows:(1)In order to solve the problem of high computational complexity of transmission map refining in the dark channel prior algorithm,this paper proposed a novel fast raw transmission map refining method based on the MMWF filter.Our improved algorithm with low time complexity,greatly improved the speed of the method,and maintained the image edge information,it suppressed the halo effect effectively,at the same time,reduced the time complexity of the algorithm.(2)A CPU plus GPU heterogeneous implementation for the parallelization of the dark channel prior video defogging was designed.CPU was mainly responsible for algorithm logic control,while GPU was mainly responsible for the data processing section.Then,we used the CUDA platform to realize the parallelization of each step of the improved algorithm and mainly solved the parallelization implementation problem of parellel planning and image convolution.Through the experimental test,the speed of GPU parallelization of the defogging can be increased by 30 to 40 times per second,and the parallelization of defogging can handle about 20 frames of the picture,basically meet the requirements of real-time..(3)Through the research of time consistent video defogging algorithm,the steps of large computation and high complexity of the algorithm was parallelized on the GPU side by using CUDA programming model and memory model,and used instruction set and program code optimization,shared memory allocation,to achieve the purpose of optimization,so as to further accelerate the speed of image processing.Finally,the results of CUDA platform parallelization was displayed by MFC interface,GPU paralleelization speed up 3-4 times.