Research On Real-Time Defogging Algorithm Of Dark Channel Based On Heterogeneous Platform

With the rapid development of science and technology,image processing and media display have received great attention.There is an increasing demand for high definition multimedia content,especially the establishment and processing of real-time HD video and image.Artificial intelligence,machine recognition,video monitoring and unmanned vehicle all need real-time HD video as the input source,especially in the harsh environment.However,the current single CPU cannot handle a large number of high-definition data contained in the multimedia content,making higher performance computing capabilities to undertake analysis and processing in an urgent need.With the corntinuous improvement of GPU performance,the CPU+GPU heterogeneous computing environment and the OPENCL heterogeneous parallel computing are rising gradually.The parallel computing power of GPU effectively improves the computer's capacity for large-scale processing.In this paper,the research and optimization of high definition real-time video defogging algorithm,as well as its engineering application under heterogeneous computing environment,will be elaborated below.(1)At present,algorithm complexity and image defog effect are the focus of fog removal algorithm.Few studies have shed light on speeding up defogging rate by parallel acceleration.High definition video has higher resolution than ordinary videos,so it is more difficult to deal with HD video in real time.In this paper,a parallel defogging algorithm based on theory of dark channel prior is proposed.In order to achieve the effect of comparsion,we first implement the single CPU serial version of the algorithm,and conduct the corresponding CPU+GPU heterogeneous parallel version according to the single CPU version.Based on the heterogeneous computing environment,GPU parallel computing cap.ability to undertake the complicated task can be made full use of.By using the memory hierarchy of AMD GPU,the memory access delay is reduced while the algorithm parallelism can be improved,and the execution time of the dark channel prior algorithm will be shortened.Finally,we transplant an independent parallel module in the open source multimedia framework(FFMPEG)to carry out the experiment of the effect of defogging.The experimental results show that the time of processing HD video is 236ms,with a good effect of defogging,and the algorithm performs,at least two times faster than before optimization.(2)The existing image degradation physical model in fog is not applicable when it contains the sky area.The conventional method of defogging has some problems,such as color distortion,scene edge halo effect and so on.This paper analyzes the model features of the sky and combined with the detection of the sky region proposing,a defogging method based on sky segmentation.First,we use MATLAB to draw a large number of sky image histogram and contour map of brightness.The analysis of luminance images shows that the sky area has a brightness value and has a certain continuity and smoothness.Secondly,combine the adaptive threshold algorithm with the sky luminance feature to determine the segmentation threshold.Non sky area and sky area are treated with different defog methods.Finally,use gamma correction and automatic contrast enhancement to deal with the dark problem.The result reveal that the sky segmentation method can be used to identify the sky area effectively,the image after processing is more natural and smooth,and there is no obvious distortion.(3)The software platform of defogging is realized by applying the multimedia development tool(FFMPEG),the video playback window(SDL)and Qt,the software platform meets the functions of video/image source input,split-screen play,going forward,moving backward,image preservation,and etc.