Efficient Parallel Implementation Of Key Algorithms In Bayer Pattern Image Signal Processing

With the popularity of digital devices and embedded devices,the requirement of the speed and quality of image processing are getting higher and higher.Graphic Processing Unit(GPU)becomes more and more popular because of its low price and high performance.Because GPU has multiple computing units,so how to use this characteristics to accelerate image processing becomes a research hotspot in industry and academia.This thesis introduces how to use OpenCL language to modify five key image signal processing algorithms including demosaicing algorithm,dead pixel detection and correction algorithm,denoising algorithm,white balance algorithm and histogram equalization algorithm when running on a GPU based Bayer-Pattern image.And this thesis further introduces how to optimize according to the GPU architecture.Experimental research is carried out under different image sizes and different types of GPU.Compared with the CPU version,the speed of optimized OpenCL demosaicing algorithm is increased by about 200 times,the speed of dead pixel detection and correction algorithm is increased by about 80 times,white balance algorithm is increased by about 2.5 times,histogram equalization algorithm is increased by about 3 times.Especially for more complex denoising algorithm,the speed is increased by about 149 times.The experimental results show that the image processing algorithm can achieve good acceleration effect through OpenCL.The innovation of this topic is that the algorithm is not simply and directly implemented by OpenCL programming.Combining with the characteristics of GPU architecture,we further modify the algorithm from the aspects of Bandwidth utilization,data access frequency,and usage of computing units to achieve good results.