Research On Image Compressed Sensing Parallel Reconstruction Technology

Signal sparse representation、measurement matrix and signal reconstruction are the main threeparts of compressed sensing technology. Reconstruction is a kind of signal recovery process fromthe observation data by solving a highly nonlinear problem. The whole process is so timeconsuming that cannot be widely applied to the practice. This thesis takes the GPU parallelprocessing technolog for signal reconstruction to shorten the signal reconstruction time on thepremise of no loss of quality of reconstructed signals.Date parallel processing is to put the data mapping to the threads which run independently eachother. GPU parallel computing engine can solve complex computing tasks more efficiently than theCPU, it has more computing cell. This paper completed the following three aspects using thistechnology.First this thesis researches a parallel signal reconstruction method about whole imagecompressed sensing based on GPU. Analysised the main time-consuming steps of OMP, thenimplementated parallel matrix multiplication using a kind of block algorithm. To speedup the matrixinversion module, the matix-inverse-update algorithm based on GPU is adopted. In this wayrealized the parallel optimization of OMP and its improved algorithm ROMP and CoSaMP.Secondly, this thesis researches a multiple granularity parallel algorithm, designs a parallelsignal reconstruction in block level.The CPU is responsible for the process control and a small partsof computing. The GPU is mainly responsible for the computing tasks. This structure can be furtheroptimized the signal reconstruction.Finaly, this thesis implements a compressed sensing optimization scheme based on the GeneticAlgorithm. This scheme introduces Elite Reserve strategy, it can shorten the signal reconstructiontime and ensure that can get global optimal solution. The next step in this thesis, we implement theparallel optimization of this scheme on GPU. The result is greatly reduced signal reconstructiontime.