Research On Techniques Of Image Quality Improvement Based On Compressed Sensing And Sparse Tensor Representation

Under the guiding ideology of the “Land-Sea Coordination” proposed by 19 th National Congress of the Communist Party of China and the “Green Development” put forward by the 13 th Five-Year Plan,the exploration of ocean information and the monitoring of environmental pollution are the main battlefields of scientific and technological innovation,and the acquisition and processing of relevant images and videos are important parts of them.Facing the dark deep water and the colorless pollution,ordinary imaging equipments cannot image the detected targets,but imaging sonar and imaging spectrometer can complete the task of visualizing them.Therefore,it is of great significance to study sonar images and hyperspectral images.Due to the limitation of imaging principle,there are serious degradation problems in sonar images and hyperspectral images respectively: speckle noise pollution caused by echo interference and low spatial resolution caused by imaging spectrometer to ensure photon energy.In view of the above problem,based on Compressed Sensing(CS)and tensor representation theory,sonar images and hyperspectral images are selected as research objects to detect “invisible”information in this dissertation,and image quality improvement is carried out from two aspects of noise suppression and super-resolution.In addition,the research is extended to corresponding video signals according to the increasing sequence of signal dimensions.The main research work includes the following contents:Firstly,aiming at the problem that most of the existing denoising algorithms of sonar images ignore the sparsity,based on sonar multiplicative noise model,this dissertation proposes a CS-based speckle noise suppression method.At the beginning,in the process of sparse representation,sonar images are represented by over-complete dictionaries,and an Orthogonal Matching Pursuit algorithm for Sonar Despeckling(SDOMP)is proposed to update the sparse coefficient.Then,in the process of updating the dictionary,a K Singular Value Decomposition algorithm for Sonar Despeckling(SDK-SVD)is proposed,and the denominator of the constraint item is used as the weight function of singular value decomposition to update the dictionary atoms.Finally,in the process of image reconstruction,the updated dictionary and the sparse coefficients are used to obtain the sonar images after denoising.In the experiments,the proposed method is compared with existing mainstream speckle noise suppression algorithms on the test images and the real sonar images respectively.The results show that the algorithm proposed in this dissertation is superior to other algorithms in five quantitative evaluation indexes and has better despeckling effect in subjective visual perception comparison.Secondly,aiming at the problem of making effective use of inter-frame information in the process of sonar video speckle noise suppression,this dissertation extends the work of previous research content in the time dimension,introduces the tensor model,and proposes a speckle noise suppression algorithm based on compressed sensing and tensor model.First of all,tensor model is introduced in the sparse representation stage to get the“integrated” description of sonar video,by using the rank-1 tensor dictionary to represent the video blocks,a CP tensor decomposition based OMP(CP-OMP)method is proposed for sparse coding of video blocks.After that,in the dictionary updating stage,CP tensor decomposition is adopted instead of the traditional K-SVD to update the dictionary.Finally,the denoised sonar video is obtained by updated tensor dictionary and sparse coefficient vectors.In the experimental part,the proposed method in this dissertation is compared with the existing mainstream speckle noise suppression algorithm on the real sonar video signal.In terms of the objective variance of residual speckle noise and subjective visual perception,the proposed algorithm has achieved good denoising effects.Thirdly,aiming at the dependence of the existing Super-resolution(SR)algorithm on the high spatial resolution image(e.g.panchromatic image,etc.),this dissertation proposes a hyperspectral image SR algorithm based on CS and sequential information.At first,based on CS theory,low-resolution image sequences are used to train low-resolution dictionaries and sparse coefficients.Then,Projection onto Convex Sets(POCS)algorithm is used to fuse low-resolution dictionary atoms to get the high-resolution dictionary.Finally,highresolution image is obtained by using high-resolution dictionary and sparse coefficients.In the experimental part,the proposed algorithm is compared with the state-of-art algorithms on the test image and the real HSI.The results show that the proposed algorithm performs better at Information Entropy(IE)and other four subjective,and also objective indexes.On this basis,aiming at the problem of introducing redundant information in the super-resolution process,a method of region of interest extraction based on tensor block reordering is proposed.The experiment shows that this algorithm has better performance than the traditional Zig-zag scanning sorting method.Fourthly,aiming at the problem of high dimensional representation and low spatial resolution of hyperspectral video,this dissertation expands the work of the third research content in the time dimension,introduces the fourth-order tensor model,and proposes a hyperspectral video super-resolution algorithm based on CS and cumulative tensor decomposition model.At first,hyperspectral video is described using the fourth-order tensor,and a key frame detection algorithm based on Cumulative Tensor CP Factorization(CTCF)is proposed.After that,based on CS theory,a super-resolution algorithm based on Sparse Tensor Tucker Factorization(STTF)is proposed.Sparse Tensor representation is used to decompose hyperspectral video frames,and dictionaries are trained in three dimensions.At last,three dictionaries and core tensors are used to calculate the highresolution video frames.In this dissertation,comparative experiments are conducted on actual hyperspectral video data.The results show that,among the comparison of five subjective and objective evaluation indicators,such as the ROC(Receiver Operating Characteristic)curve,the algorithm proposed in this dissertation is superior to the mainstream algorithm in terms of detection and enhancement of hyperspectral video.Through the above researches,this dissertation applies the theory of CS and the tensor representation model into two important image quality improvement methods for detecting "invisible" information,and further completes the quality improvement of corresponding video signals by using the time dimension information.The methods proposed in this dissertation not only improves the performance of the image quality improvement,but also extends the application range and increases the practicability of the algorithms.