Research On Heterogeneous Acceleration Technique For High Dynamic Image Fusion

The dynamic range of real world scenes can be as high as more than 10 orders of magnitude,but the generally shooting and displaying devices often have lower dynamic and can't match the dynamic of real world scenes.The resulting single image still has an overexposed bright region and an exposed dark area,which can't fully represent the true information of the scene,even if the exposure is adjusted by adjusting the aperture and shutter speed.More than this,images will be affected by noise at any time during shooting and stocking,which will further influence the expression of the details of the image.At the same time,with the increase of the image resolution,the time required for image de-noising and multi exposure fusion can be as long as several seconds.These reasons restrict the further promotion and application of the algorithm.Therefore,it is very important to study the acceleration method for high dynamic range imaging in the present of noise.Aiming at the problem of noise,the image de-noising method is presented in this paper;according to the detail loss problem caused by insufficient of dynamic range of shooting and displaying equipment,high dynamic range image fusion technology is presented in this paper;aiming at the problems of slow processing speed of the algorithm,the CPU + GPU heterogeneous parallel architecture is presented in this paper to accelerate the algorithm.Firstly,spatial filtering for images with different exposure of the same scene to suppress the noise,and then,getting the weight map whose weight factor is saturation and contrast in the HSI color space,then refusing the brightness map according to the weight map under different resolutions on the principle of pyramid.Finally,the result is more close to the real scene,the color is vivid and the information is kept intact.In order to speed up the processing speed,this paper mainly study the algorithm acceleration from two aspects: data access and data computation.In the aspect of data access optimization,we can shorten the access time by reusing data from cache to local memory,cache fixed template to constant memory and merging kernel function.In the aspect of data computation,the method of coordinate transformation is used to avoid duplication of calculations,the use of look-up tables instead of the calculation,a work item deal with multiple pixels and other optimization methods to reduce the amount of calculation.This paper uses bilateral filtering and non-local means of multi exposure image de-noising,the experimental results show that under different scenarios,two algorithms can obviously improve the image quality of multi exposure fusion results.Compared with the non-direct fusion de-noising results,three scenarios the peak signal-to-noise ratio improved by 29.77% on average and the average structural similarity up to 10.85% when application of bilateral filtering;three scenarios the peak signal-to-noise ratio improved by 32.21% on average and the average structural similarity up to 12.7% when application of non-local average filtering.Three scenarios through multiple exposure fusion,compared with image quality before fusion in image sequences with the highest average gradient were increased by 9.70%,31.02%,18.60%,the spatial frequency was increased by 11.99%,18.25%,18.49%.Analysis and comparison of the CPU + GPU heterogeneous parallel architecture to accelerate the running time and the acceleration of the running time,the three scenarios in this article,respectively,were 57.3 times,71.1 times and 42.8 times faster than.The results of this study show that the speed of noise removal and processing has a positive effect on the further application of high dynamic image multi exposure fusion.