Paralel görüntü onarımı

In this thesis, we are concerned with the image restoration problem which hasbeen formulated in the literature as a system of linear inequalities. With this for-mulation, the resulting constraint matrix is an unstructured sparse-matrix andeven with small size images we end up with huge matrices. So, to solve therestoration problem, we have used the surrogate constraint methods, that canwork efficiently for large size problems and are amenable for parallel implemen-tations. Among the surrogate constraint methods, the basic method considers allof the violated constraints in the system and performs a single block projectionin each step. On the other hand, parallel method considers a subset of the con-straints, and makes simultaneous block projections. Using several partitioningstrategies and adopting different communication models we have realized severalparallel implementations of the two methods. We have used the hypergraph par-titioning based decomposition methods in order to minimize the communicationcosts while ensuring load balance among the processors. The implementationsare evaluated based on the per iteration performance and on the overall perfor-mance. Besides, the effects of different partitioning strategies on the speed ofconvergence are investigated. The experimental results reveal that the proposedparallelization schemes have practical usage in the restoration problem and inmany other real-world applications which can be modeled as a system of linearinequalities.Keywords: Parallel image restoration, distortion, parallel algorithms, linear feasi-bility, surrogate constraint method, hypergraph partitioning, rowwise partition-ing, checkerboard partitioning, fine-grain partitioning, point-to-point communi-cation, all-to-all communication, convergence rate.