Real-time Super-resolution And Stereoscopic View Genera-tion With GPU/Multicore CPU Based Parallel Computing

In recent years, much has been made of the computing industry’s widespread shiftto parallel computing. Driven by the insatiable market demand for realtime,high-definition3D graphics, the programmable Graphic Processor Unit(GPU) hasevolved into a highly parallel, multithreaded, manycore processor with tremendouscomputational horsepower and very high memory bandwidth. The rapid development ofGPU on performance gradually makes it a hot research topic that uses GPU toaccelerate general-purpose computing. Up to now, a variety of industries andapplication areas have enjoyed a great deal of success including orders-of-magnitudeperformance improvement by choosing to build their parallel computation on GPU. Onthe other hand, traditional Central Processing Unit(CPU) has also developed into thedirection of multi-core parallel architecture which providing us with considerablecomputing power that can ease our burden for computational complexity.In this thesis, we focus on two image/video processing problems: super-resolutionreconstruction and view synthesis for stereoscopic content generation. First of all, weare committed to obtain high quality processing results by proposing effective methods.Then the real-time performance is achieved by getting the speed-up benefit withGPU-accelerated parallel computation. We also implement our super-resolutionalgorithm on a multicore-based platform to balance the computing complexity andresult quality for a further study about the parallel computing. The main achievementsin our work are described as follows:1．In image/video super-resolution reconstruction processing, we design a simplebut effective framework on GPU for both image and video super-resolution and a highlyparallelizable algorithm. The realization of the system is optimized to achieve real-timeperformance by considering the computational bottleneck of our method.2．Our super-resolution approach is implemented on multicore CPU platformparallelly by mapping each sub-problem to an independent processor core and theblocking artifacts caused by partitioning is successfully avoided with an effectivedeblocking method.3．In dealing with stereoscopic view generation, we propose a depth adaptivepreprocessing of depth map to separately handle the depth values of foreground andbackground so that the geometric distortions and losses in depth cues could be reducedreasonably. The proposed method also integrates the depth adaptive technique into theframework of hierarchal hole-filling to improve the rendering quality of final results. In addition, super-resolution technology is utilized in our hierarchal hole-filling procedureto reduce the blurred effects occured in the hole places.4．The proposed stereoscopic view generation system is optimally implemented onGPU to achieve real-time performance based on the mainly improvements of memoryaccess efficiency and parallel execution performance.Experimental results demonstrate that both the proposed methods achieve superiorperformance in comparison with the existing methods with respect to result quality andexecuting time.This work is supported by the National Natural Science Foundation of China (No.61271298).