The Research And Implementation Of Real-time Video Dehazing Based On CPU-GPU Heterogeneous Platform

The images captured in outdoor haze weather have poor sharpness,low contrast,and loss of visual information.It is difficult to extract the useful information from the images.Degraded images cause certain difficulties in applications such as monitoring and target detection.In order to obtain clear images,dehazing research came into being.With the development of digital image technology and hardware equipment,video processing has become a focus in daily life.In order to obtain clear video,two problems need to be solved: adaptive dehazing with good results,and the efficiency of the algorithm must meet the real-time requirements.Based on the dark channel prior dehazing algorithm,this paper completes real-time dehazing research on a CPU-GPU heterogeneous platform.The main work of this paper is as follows:(1)Optimization of dehazing algorithm.In this paper,the morphological expansion operation is used to eliminate the halo phenomenon in the image where the depth of field changes suddenly.Because the sky area in the image does not meet the dark channel prior theory,this will generate noise in the dehazed image.Therefore,this paper first determines whether the image contains the sky area,then compensates the transmittance for the sky area,stretches the transmittance of the sky part and eliminates the noise.We uses single image haze removal algorithm to do video dehazing.Due to the severe jump of the atmospheric light estimates in adjacent frames,flicker problems is prone to occur in the recovered video.This paper uses the inter-frame relationship of the video to constrain the atmospheric light estimates.We must not only ensure that the atmospheric light value automatically adjusts with the change of the scene,but also ensure that the atmospheric light value does not jump frequently.Through a series of optimizations to the algorithm,the quality of the dehazed image can be ensured.(2)Parallel algorithm optimization.In this paper,the dehazing algorithm involves four types of operations: minimum filtering,mean filtering,atmospheric light value estimation,and ordinary matrix calculation.The minimum filtering algorithm uses a parallelized fast minimum filtering algorithm for depth optimization.The mean filtering algorithm,based on the idea of integral graph,gradually improves the optimization efficiency from sliding window algorithm to parallel scanning algorithm.For the estimation of atmospheric light,this paper proposes a vertical block algorithm,which uses a reduction algorithm to process the data in the block.For ordinary matrix calculation,this paper uses a single thread to calculate a single-point pixel.On the algorithm level,a deep parallel optimization of the dehazing scheme is made.(3)Data access optimization.In the kernel functions,there are some data that need to be accessed frequently.This paper uses the shared memory of the GPU to speed up data access.In the filtering algorithm,this paper adopts the method of row and column separation.In order to speed up the access to longitudinal data in global memory,this paper uses the transposition method to make the longitudinal data meet the coalesced memory access.In video image processing,image data needs to be frequently transmitted between the two platforms.In order to reduce the transmission delay,this paper uses an asynchronous memory transfer method to hide the data transmission time between the platforms.This paper optimizes the dehazing strategy from the aspects of algorithm and memory,and implements it on the hardware system.The solution in this paper greatly improves the quality of the recovered image and improves the speed of the algorithm.On the GTX 1080 Ti platform,the speed up ratio of 1080 P video reaches 480 times,and the single frame processing time is less than 7ms.We tested 1080 P video on multiple platforms,and the results all meet the requirements of real-time dehazing.