Self-adaptive Compressive Sensing Of Nature Image

Compressive sensing (CS) is an emerging approach for acquisition of signals. For two-dimensional images, the compressive sensing based on image blocks is always used to reduce size of sampling matrix and the complexity of the sampling and reconstruction. While the block CS literature has assumed that all the blocks have the identical sparsity or compressibility which cause the same sampling ratio is assigned to each block. However, the image blocks which contain different structure information and possess different sparsity and compressibility are too different to share the same sampling ratio. In this paper, we propose two frameworks for self-adaptive sampling strategy based on variance and total variation. Each block will be given a proper sampling ratio on the basis of its sparcity. And at the stage of reconstruction, to improve the quality of blocks, especially the smooth blocks, non-local total variation (NLTV) model is adopted to reconstruct the original image block-wisely with some non local prior information, and augmented lagrangian method (ALM) is used to search the optimal solution of NLTV minimization problem.In fact, compared with the l0norm minimum prior, many more effective prior information can be used to reconstruct the original image, as NLTV, KSVD and BM3D such famous denoising methods. We also designed a robust sensing matrix by learning a mass of image blocks using principal component analysis (PCA). A proof has been given that most energy of the image block can be extracted by the trained sensing matrix.Some experiment on real data and results shows both the proposed framework above can capture the structure information of the image, and reconstruct the original image under the low average sampling ratio but with a perfect quality, especially the detail or texture information. A thoroughly analysis are taken to validate the superiority of our proposed method to traditional framework.