Implementing Parallel Computation For Image Restoration Via Total Variation Model On GPU

In the process of physical imaging and transmission,due to imaging systems and transmission media and other incomplete lead to the quality of the image decreased,this phenomenon is called image degradation.Image restoration has been a hot topic in the field of image processing because image degradation leads to difficulties in post-processing such as image segmentation,image feature extraction and target recognition.At present,the rapid development of GPU is much higher than the speed of Moore's Law.The single precision floating-point processing capability of the main-stream GPU and the external memory bandwidth have a significant advantage over the same period of CPU and in the cost and power consumption do not have to pay too much price.NVIDIA'S CUDA unified computing device architecture based on the G-PU,in programming,optimization and other aspects has been significantly improved and greatly enhanced the GPU's general computing power which provides new solu-tions for large-scale intensive computing.Blur and noise are typical image degradation.Image restoration is to remove the noise and blur,while preserve the edge of the image,texture and other details as much as possible,which lays the foundation of machine vision,machine learning,machine identification.We can obtain a higher quality image by regularization methodology based on total variation,because this model is highly effective in removing noise,blur,or other unwanted fine-scale detail,while preserving edges.When the image pixels are getting bigger and bigger and the resolution is getting higher and higher,the calculat-ed amount of total variation model increases exponentially,which brings challenge in image real-time processing.Aiming at this problem,this thesis explores the parallel implementation of the total variational dual model of image restoration algorithm on the GPU device to improve the computational efficiency.This thesis introduces GPU programming knowledge and total variational dual model in detail,using the gradient projection algorithm to solve total variational dual model.The algorithm is implement-ed in parallel computation on the GPU and compares it with the efficiency of the imple-mented on the CPU.The experimental results show that the efficiency of GPU parallel execution is higher than that of CPU.With the increase of image size,the advantages of GPU parallel computing are more obvious.