Image Reconstruction Algorithm Based On Dynamic Dictionary Learning

The effective compressed sensing of the image depends on the prior knowledge of the appropriate dictionary for sparse representation of the target image,which is often lacking in the actual natural image reconstruction.The traditional solutions are as follows:taking the off-the-shelf over-complete dictionary to reconstruct the image,which is usually pre-trained by a large number of training sets,but is not necessarily suitable for any image to be reconstructed,and overfitting is the primary obstacle,that is,the model generalization results in the loss of the entire image.In order to solve the problem that the theory of compressed sensing has too high a priori requirement on the signal,the blind compressed sensing theory comes into being,which is to train the dictionary to reconstruct the image without prior knowledge.However,such a method requires very high structure of dictionary and sparse coefficient,and the dictionary sparse expression obtained by the method has poor generalization ability.Feature transfer is introduced into dictionary learning to solve the above problems.An online dynamic dictionary learning algorithm based on transfer learning is proposed,which uses sample subset to calculate the target error and gradient,adjusting the atoms related to the target domain in the pre-learning dictionary online,and removing the atoms unrelated to the target domain,so as to improve the sparse representation efficiency of the dictionary.The principle and application of image reconstruction algorithm based on compressed sensing are introduced in detail from three aspects:sparse representation,measurement matrix and incoherent perception,sparse reconstruction.The ODDML algorithm adopts the strategies of transfer learning and dynamic gradient calculation to reduce the number of training samples and improve the sparse expression accuracy of target images.The cross-validation strategy is adopted to reduce the dynamic gradient computation and facilitate the online learning of dictionaries.The effects of cross validation and dictionary transfer learning strategies on the performance of the proposed ODDML reconstruction are analyzed and compared with other mainstream dictionary learning algorithms in the application of image reconstruction.The simulation results show that the ODDML algorithm has higher convergence speed,higher PSNR and higher structure similarity.