Research On Characters Correction Methods Based On Low-rank And Sparse Decomposition

With the advance of modern science and technology, the improvement of the computerperformance and the further study of researchers, application research of patternrecognition problem has achieved remarkable results. For example, distinguishing differentcharacter image by face recognition, separating different mood through the differentexpressions, etc. In the process of application study, character recognition, voicerecognition, face recognition has become several prominent aspects in the field of patternrecognition research. As a relatively old field of research in pattern Recognition, OCR(Optical Character Recognition)has important position in the history of the patternrecognition.To further enhance the accuracy of online OCR recognition, and taking intoaccount the error noise and occlusion interference in character image, using low-rank andsparse image recovery method for image denoising and the correction is a feasible method.Low-rank matrix recovery problems is an important data analysis tool,which derived fromthe very popular compressed sensing technology in recent years, and has been widely usedin computer vision, image processing, recommendation systems, text analysis.Based on a careful study of the status quo at home and abroad this paper conduct acomprehensive analysis and summary for the existing algorithms and applications oflow-rank matrix recovery problem and points out the shortages of the them. Low-rankmatrix recovery method has existed some defects, such as large amount of calculation, thematrix to handle is too smaller, which making this method in many cases cannot fully playto their advantages.For the lack of Augmented Lagrange multipliers (IALM) convergence, this paperproposes an improved algorithm based on augmented lagrange multiplier method: at eachiteration process, a new iteration steps use some amendments to correct this predictor,which is generated by theALM output.Although there have been augmented lagrange multiplier method proposed, but thelow-rank recovery algorithm problem has still very slow efficiency and a large amount ofcomputation. This paper researches a new parallel splitting augmented lagrange multipliermethod (PSALM), which combines the parallel separation method with theALM algorithm,then form a new iteration through a convex combination step, which can guaranteeconvergence at the same time to improve the computing speed of the algorithm.This article use parallel separation augmented Lagrangian multiplier method to solvethe character correct problem of the OCR. Compared with ALM and improved ALMalgorithm that has been proposed in the literature, the proposed method can ensure thecorrect convergence, while the efficiency of the algorithm has great improvement. Underthe conditions that the current computer and portable terminal hardware is CUDA platform of multi-CPU and GPU, the application of multi-core processing and the implementation ofparallel ALM approach is particularly meaningful and operability.