Design Technique Investigating On Remote Sensing Image Restoration And Super-Resolution Parallel Processing System

The remote sensing image has been affected by blur, noise and cloud in its imaging process, so that it suffered from degradations of quality, detail and resolution. At the same time, the image has also been down-sampled by the A/D transformation, leading to frequency aliasing, spectrum distortion and further reduction in resolution. The main task of image restoration and super-resolution algorithms is to eliminate above degradations and increase the image's clarity, contrast and resolution. This thesis tried to improve resolution of the degraded remote sensing image effectively through innovative researches on parallel system design and its algorithms realization of remote sensing image restoration and super-resolution, without changing the imaging system.Contents of this dissertation can be mainly divided into two parts, hardware and software. The first part includes parallel architecture selection and hardware sytem implementation, the second part includes establishment of parallel algorithm model, parallel algorithm design and its performance optimization.Parallel technology has been widely used in various areas of image processing. However, there are still a lot of problems unresolved preventing its application and development, just as large differences in the underlying hardware system, high cost, non-uniform algorithm standards, poor portability and so on. Research on the architectures of image parallel processing systems is the basis of system design. It plays an important role in achieving optimal conversion from algorithm to structure. In this paper, we firstly proposed a classification of image parallel processing systems, analyzing and comparing different parallel architectures. It can be conclusion that the development trend of image parallel processing system is generalization and construction integration. Then research on the two most generic parallel image processing architectures of DSP and Cluster was executed particularly. We performed image restoration and super-resolution algorithms on each of them. The experiment results show that DSP is adapted to fast response of small images and simple algorithms. On the contrary, Cluster is good at real-time processing of vast volumes images and complex algorithms. If our system is designed by DSP, its scale is too large to implement and manage, so we choiced the Cluster architecture.Construction of the image restoration and super-resolution parallel system is depended on its architecture and the characters of image restoration and super-resolution algorithms. In this paper, system established works as a SMP Cluster that mainly consist of fat node, Infiniband switches and fiber disk arrays. This structure is consistent with the hybrid programming model based on OpenMP and MPI, combining advantages of fine-grained parallelism and coarse-grained parallelism. It is a high performance system. By further studies, its performance model and expression was presented. This model demonstrates the relationship among processors'number, system speedup and efficiency. Its analysis shows that there is a limit or optimal value of the processors'number. When the value is exceeded, contribution of the new added processors will be covered with increase communication overhead, so system performance decreases instead.Research status of parallel technology is development of software far behind that of hardware. So research on the image restoration and super-resolution parallel algorithm implementation techniques is significant. A model of communication costs was established through studies of the key factors affected the parallel system performance, just as load, communication and I/O. Inherent communication, additional communication, cost, delay, and conflict were deeply discussed according to this model. It can be found that merger communication is a good communication optimization strategy. Then an abstract hierarchical structure model of I/O system was also established and its optimization strategies were given. Finally, an image restoration and super-resolution parallel algorithm model of the PPCTS structure combining with data decomposition and pipeline parallel techniques was proposed.Remote sensing image restoration and super-resolution parallel algorithms were established based on the domain decomposition method and the above PPCTS parallel algorithm model. We proposed a second-order PDE-based parallel diffusion denoise algorithm for image restoration and a frequency compensation and expansion parallel algorithm for image super-resolution. The second-order PDE-based parallel diffusion denoise algorithm was proposed based on the research of PDE denoising theories and parallel PDE techniques. It was mainly used to remove the Gaussian white noise and the Poisson noise. Experimental results show that this new algorithm can remove noises efficiently while protecting the edge of images. Its processing effect is satisfied and speed is fast, meeting the real-time application needs. The frequency compensation and expansion parallel algorithm was designed through a combination of a frequency interpolation and enhancement algorithm, a frequency compensation filter and parallelization of FFT and matrix multiplication algorithms. It doesn't only eliminate spectrum distortion, but also separates the high frequency components from the low frequency components and supplement the lost high frequency components. So it extends frequency spectrum and improves frequency structure and enhances resolution effectively. The processing speed of it is very high too. Experimental results on a parallel system with four processors show that the two parallel algorithms all achieve good speedup, efficiency, scalability and portability. Their speedups are more than 3 times and efficiencies exceed 75%, up to 92.9%.