The Research Of Real-time Image Dehazing Based On Dark Channel Prior

Due to the influence of hazy weather,outdoor videos and images are inevitably degraded,resulting in a serious information loss.This leads to certain difficulties in such areas as outdoor target recognition and tracking,and road visual surveillance.Therefore,image dehazing has broad application prospects.The development of image dehazing algorithm has achieved good results.For example,single image haze removal based on dark channel prior,which is widely used in the field of computer vision due to excellent performance and lower computational complexity.However,with the development of high-definition video(HD video)in recent years,real-time image dehazing is becoming a big challenge.The dark channel algorithm is difficult to real-time process on CPU-based platform,which limits its further promotion.Currently,existing research schemes using GPU to accelerate dehazing can only achieve real-time processing of lower resolution images,and there is still a considerable distance from the goal of real-time dehazing for HD images.Based on the study of dark channel prior,we propose a GPU-accelerated parallel computing implementation,which realizes the realtime dehazing for HD images.The main innovative work of this thesis is as follows:(1)The optimization of the dark channel prior application in video dehazing.For the frameby-frame processing of dark channel prior applied into the video dehazing,a prior-prediction method for calculating the atmospheric light is proposed,which effectively controls the video flicker effect caused by the different atmospheric light values of adjacent frames.The improved algorithm can be well applied to the video dehazing,and obtains a better performance.(2)A high performance GPU-accelerated parallel computing implementation is proposed and achieve real-time dehazing for HD images.For several modules with high computational complexity,a parallel algorithm based on transposed integral imgae is proposed to realize the mean filter.The efficiency of guided filter is further improved by incorporating steps.A fast filtering algorithm is introduced to optimize the minimum filter.In order to ensure the efficiency of the algorithm,a simplified block-based atmospheric light value calculation method is proposed in combination with the reduction algorithm.Experiment results indicate that,the algorithm can perform the 1080 p dehazing with 40 frames per second and 52 frames per second on Tesla C2075 and K80 GPU,respectively without significant accuracy loss.And the speedup up to 31 x and 40 x,respectively,can meet the requirements of real-time HD image dehazing.