Research Of Image Inpainting Technology Based On Compressed Sensing

Image inpainting is to reconstruct the lost or deteriorated parts of imagesby using the rest parts of information. As a novel signal processing technique,the compressed sensing (CS) theory has proved that the source information canbe fully reconstructed by taking advantage of the sparsity of the signals, eventhough the number of data would be deemed insufficient by the Nyquistcriterion. The sparse representation and the signal reconstruction technologiesin CS theory have introduced a novel way to solve the image inpaintingproblem.Currently, most existing inpainting methods take advantage of either thestructure information or the texture information of an image, while few workstake both of the parts into consideration. Thus some useful information isignored during the inpainting process. To solve this problem, we propose akind of CS based inpainting solution with morphological component analysis(MCA) decomposition. And two implementation methods (namely theasynchronous method and the synchronous one) of this solution are put indetail.The asynchronous method first decomposes the target image into structurepart and texture part. Then the two parts are recovered in different ways. Dueto the smoothness of the structure part and the similarity of the texture part, they are reconstructed using the total variation (TV) method in wavelet domainand the Bayesian weight method, respectively. Finally, the recovered imagecan be got by adding the two restored parts together. The experimentscompared with the TV method, the MCA-based synchronous method, theexpectation maximum method and the Bayesian weight method prove that theperformance of three proposed method is better than the existing commonlyused methods.Besides, as the asynchronous method takes a long time during inpainting,an improved scheme based on MCA synchronous inpainting is proposed. Alearning based over-complete dictionary is introduced in this improved schemeto refine the recovery of the texture part. The learning based dictionary used inthis method is K-Singular Value Decomposition (K-SVD). The adaptabilityand the recover performance of this scheme are also proved by a seriesexperiments.