Key Technology Research On Composition And Restoration Of Image

High precision digital collection of murals and the digital mural repairing are crucial technologies to murals permanent protection. In this thesis we studies two important problems of Dunhuang murals composition and restoration, i. e. , the composition of large image being very slow and memory demanding, and the noise reduction being difficulty due to the diverse causes of noise. Our work consists the efforts as below:(1) In order to tackle the problem of slow performance and big memory requirement of large image composition, we propose a fast image composition method for large murals based on uniform sparse sampling.Firstly, our algorithm divides an original source image into many quadrilateral patches, and discards inner pixels and reserves the border pixels to constitute two low resolution images (Border-images) on which the image compositing operator is implemented. Finally according to the solution of Border-images the result image is evaluated by concurrently computing inner pixels of every quadrilateral patch on GPU. The evaluation of these patches are independent to each other so that our algorithm has good parallelism. Compared with the currently available methods, this algorithm can rapidly process very large image composition with relatively limited memory space. The algorithm can run on parallel on GPUs;(2) In order to resolve the problem of low parallelism in general image composition, we propose a fast composition algorithm of murals large image with matrix tearing. This algorithm firstly divides a sparse matrix into many different matrices with overlapping blocks by geometric degradation,that each block could be processed by GPU independently on parallel. The algorithm significantly improves the operation rate with low memory consumption. In order to further address the problem of limited parallelism of image processing programs, we propose a method of automatic parallelization based on compile, which adopts an optimized speculative execution strategy by borrowing some idea from Speculative Multi-Threading (SpMT) processors. This method exploits the implicit parallelism in the program automatically without any extra development or maintenance cost, and makes better use of computing resources of multi-core processors.(3) In order to improve the murals image restoration performance which suffers from the diverse types of noise, this thesis proposes an adaptive mural image restoration algorithm based on a priori method. This algorithm combines the local and non-local variation of the image, which fits well for mural denoising and completion tasks. When further combined with sparse matrix noise reduction, this algorithm is robust when restoring murals eroded by diverse types of noise.