Multiple Exposure Fusion Algorithm Parallel Optimization Based On OpenCL

With the advent of the digital information age,people are more urgent pursuit of highdefinition lifelike viewing experience.The rise and development of high dynamic range imaging has provided digital media with a new and powerful impetus for the development of high-definition,high-information discovery.Multi-exposure image fusion provides a convenient way to synthesize high dynamic range images.Without knowing the camera's response parameters,multiple images with different exposures can be used to synthesize a clear and dark area with clear detail information.However,most of the multi-exposure fusion algorithms have a large amount of data,a high complexity of the algorithm,and a timeconsuming implementation process,which can not meet the needs of practical applications.In this thesis,using GPU's powerful parallel computing capabilities,two classic multiexposure fusion algorithms are accelerated based on Open CL parallel programming specification and GPU parallel architecture,which improves the running speed of the algorithm.This thesis first studies the parallel computing technology,analyzes and studies the popular parallel implementation methods and parallel devices,analyzes the architecture of AMD GPU,and paves the way for parallel acceleration of subsequent multiple exposure fusion algorithms.A multi-exposure fusion algorithm based on pyramid decomposition is analyzed in detail.A variety of conversion methods of RGB and HIS color space in the algorithm are studied and analyzed.A method that is conducive to parallelization and good conversion effect is selected for algorithm execution.For the Gaussian low-pass filtering in the pyramid decomposition process,the filter core reduction dimension optimization is performed first,and then the local memory of Open CL is used to further optimize the reuse of data in the filtering process.The multi-exposure fusion algorithm based on propagated filtering was studied in depth.Firstly,the in-depth analysis of the propagated filtering is carried out,then the multiexposure fusion weights are constructed based on the propagated filtering.Through the experimental analysis of the propagated filtering,the filter neighborhood of the filtering filteris optimized.For the weight construction process,there are too many intermediate variables,it is too time-consuming to start multiple kernel calculations,local memory and integrated multiple kernels are used to avoid unnecessary time overhead.Finally,in order to verify accelerated performance,the thesis validates serial and parallel code using multiple sets of images of different size.By using information entropy,average gradient,spatial frequency and PSNR,the results of serial and parallel fusion are compared and analyzed.The experimental results show that the fused image quality of the serial and parallel algorithms are not very different.Compared with the serial,the parallel algorithm has a larger speedup in time,and the speed performance of the algorithm is greatly improved.