Research And Implementation Of GPU Acceleration Technology For Image Processing

With the rise of deep learning,artificial intelligence,big data and other fields,the amount of data processed by computers has become larger and larger,and the current CPU has been unable to meet the research needs of these fields.GPU(Graphic Processing Units)was originally used to process graphics computing.With the development of technology,GPUs have more powerful functions,which have higher efficiency and better performance in largescale parallel computing.Using GPU parallel computing to accelerate data processing and improve the efficiency of the algorithm has become a hot research topic.Due to the rapid development of computer vision technology,the requirements for computer performance are getting higher and higher.Processing images through GPU acceleration technology is also the current hot research direction.The dark primary color a priori algorithm used in image defogging is time consuming and computationally intensive,making it difficult to achieve real-time dehazing of high-definition foggy images.For the problem of low efficiency of the algorithm,it has been proposed to use the GPU to accelerate the algorithm,but the acceleration device and the parallel method adopted cannot defogg the HD image in real time.Aiming at this problem,this paper takes image de-fog as an example to study the GPU acceleration strategy of image processing,in order to achieve real-time processing of high-definition image defogging.The specific research work of this paper includes:(1)Introduced the CUDA(Compute Unified Device Architecture)platform for GPU parallel acceleration computing,and analyzed the GPU parallel acceleration technology using CUDA platform.(2)Summarize the existing defogging algorithms and analyze their advantages and disadvantages.This paper analyzes the characteristics of the defogging algorithm based on the dark primary color a priori.For the problem that it can not defogg the foggy image with sky region,the method of using OTSU segmentation algorithm to calculate the atmospheric light value of the foggy sky image with sky region is proposed.The atmospheric light value is the mean value of the atmospheric light intensity in the sky region after the segmentation.This method can effectively solve the problem of fogging foggy images containing sky regions using the dark primary color prior algorithm.For the time-consuming problem in the defogging algorithm,the parallel acceleration of the dehazing algorithm using the CUDA platform is proposed.According to the related parallel acceleration technology,parallel analysis is performed on each small process of the defogging algorithm,and a suitable parallel method is selected.For the current defogging algorithm after GPU acceleration,there is a problem that some processes are not efficient,and the problem is solved.A new parallel acceleration method is proposed: the data in the Box_filter filtering in the steering filtering uses a parallel method in which the shared memory retains the intermediate result;in the calculation of the atmospheric light value,the parallel primitive method of different spans is used in the dark primary color map to improve the utilization of the thread.rate.Experimental results show that compared with the existing defogging algorithm using GPU parallel implementation,this paper uses GPU acceleration processing of the defogging algorithm to have higher operating efficiency,the sky part of the fog day image can also be handled very well.Using GPU-accelerated defogging algorithm to process 720 P of highdefinition fog day pictures,basically achieved real-time de-fog.