Research On Image Compressed Sensing Algorithm Based On Adaptive Sampling

The theory of compressed sensing breaks the limitation of traditional samplingtheorem, which can reconstruct original signal by fewer sampling points than Nyquistsampling theorem required. It achieves sampling and compressing at the same time. Itsoutstanding advantages are the reduction of sampling data, saving storage room, whichmakes design the system easy and reduces the requirement of sampling equipment. Thispaper does some researches on the problems of designing sparse dictionary, optimizingprojection matrix and improving reconstruction algorithm, this work can be summarizedas the following three aspects:Firstly, because of the block compressed sensing algorithm commonly uses the samesampling operator for all image patches in a whole image, the required number ofobservation to reconstruct accurately the patches with simple structural characteristics isless than complex patches. So the adaptive sampling is proposed. The observation vectorsare divided into smooth part and non-smooth part according to their enery characteristics.The smooth part is sampled again with permute discrete cosine transform. Experimentalresults show that better images can be obtained by using adaptive sampling with the samereconstructing algorithm compared with non-adaptive sampling.Secondly, because of the single dictionary can not sparsely represent the imagepatches of different types, a redundant directional dictionary is proposed which containsmany sub dictionaries. The coefficients representing image are sparser by using theredundant directional dictionary. Taking into account the similarity between the imagepatches, the constraints of non-local similarities and image sparse representation areintroduced into the reconstruction equation to improve the performance of the algorithm.Experimental results show that the proposed algorithm is better than existing blockcompressed sensing algorithm.Finally, according to the uncertainty of the observation matrix, how to construct anobservation matrix with certainty is studied. The object function to construct matrix isimplementated using different methods. The observation matrix suiting to the redundant directional dictionary is designed. Experimental results show that using the improvedobservation matrix with the same reconstruction algorithm can obtain betterreconstructed image.