Study Of Compressive Sensing Based Super-resolution Image Reconstruction

Image super-resolution reconstruction is a processing algorithm of improving the resolution without modifying the existed imaging equipment,in which the valuable prior information is used to estimate a high-resolution image.Since the reconstruction is an ill-posed problem,the usual methods have the non-uniqueness result and the weak capability of noise suppression,leading to the blurred edge region and false high-frequency information.As a novel sampling theory,compressive sensing samples the sparse or compressible signal(e.g.,natural images)with a low sampling rate,accurately reconstructs the original signal from the captured measurement data with a high probability,and brings a new method to achieve the reconstruction of the super-resolution image.This paper focuses on the research of the reconstruction method based on compressive sensing,the thesis introduces the algorithm of single image super-resolution,the design of the actual compressive imaging system and the method of overcoming the diffraction limit,so that the resolution of the image can be improved,the main contents of this paper include:1)A single image super-resolution reconstruction algorithm based on a redundant dictionary is proposed.The K-SVD algorithm is used to generate a redundant dictionary from the training set instead of the orthogonal sparse basis as the prior knowledge,which can increase the sparsity of the image,and effectively solve the incompatible problem between the degradation model and the compressive sampling model.On this basis,combining compressive sensing improves the resolution of the input image,and solves the ill-posed problem of the super-resolution reconstruction.Experimental results indicate that the proposed algorithm has a high accuracy,and the peak signal to noise ratio and structural similarity of reconstructed images are also high.2)In order to solve the problem of the low efficiency and adaptation resulting from the redundant dictionary,a single image super-resolution reconstruction algorithm based on clustering sub-dictionary is proposed.In the phase of learning the clustering sub-dictionary,the method based on the Gaussian mixture model is firstly used to cluster the image patches after screening,and then principal component analysis is used to generate the sub-dictionary from each set of the image patches,which not only improves the rate and adaptation of the atom,but also accelerates the speed of the dictionary learning.Compared with the algorithm based on a redundant dictionary,the steering kernel regression and the non-local means filter are used as the regularization terms to make the algorithm having a strong ability to suppress the noise and sharpen the edge region.3)A compressive imaging technique based on scrambled block Hadamard ensemble in the frequency domain is proposed.Due to the complex structure and difficult hardware implementation of random Gaussian and Bernoulli matrices,this technique treats the scrambled block Hadamard ensemble as the measurement matrix,uses the pure-phase only liquid crystal spatial light modulator to achieve the sensing operation of this matrix,and utilizes the 4F optical processing architecture to design the compressive imaging system.In the process of the object imaging,the iterative phase retrieval algorithm is firstly used to recover the complex amplitude distribution of the detector p lane;the reconstruction algorithm of compressive sensing is secondly used to estimate the initial complex amplitude distribution of the object,and obtain the initial intensity image;the clustering sub-dictionary and the gradient histogram preservation model are finally used to denoise the initial intensity image for improving the edge sharpness of the reconstructed image.4)In order to overcome the diffraction limit and achieve super-resolution imaging,a compressive super-resolution imaging technique based on a binary pure-phase annular filter is proposed.The optimization method of constraining the Strehl ratio is firstly utilized to design the binary pure-phase annular filter;then the filter and the random sampling mask are utilized to sample the object;the previous imaging method is finally used to achieve the object's super-resolution imaging.Experimental results indicate that this technique cannot only improve the resolving power of the imaging system nearly 1.243 times,but also reconstruct the high-resolution intensity image by using the measurement data accounted for 25%of total samples.