Design And Implementation Of Image Reconstruction System Based On Fast Convolutional Dictionary Learning Algorithm

With the development of neuroscience,scientists have conducted in-depth research on the signal processing mechanism of animal cerebral cortex,and found that the receptive field of primary visual cells in mammals has the properties of spatial locality,directivity and connectivity.Then scientists proposed a sparse coding model to describe these properties of primary visual cells.However,local blocks are independent of each other in sparse coding technology leading to that the basis in the dictionary obtained is the version after shifting with each other.In order to model the shift invariance,scientists proposed convolutional sparse coding(CSC).CSC is a bi-convex problem,and the prevailing solution is Alternating Direction Methods of Multipliers(ADMM).Found by experiments,the solution of CSC involves dealing with complex matrix inverse problem leading to a larger space and the amount of calculation.In order to solve this problem,this paper introduces precondition technology which can circumvent matrix inverse operations,reduce the computational complexity and improve convergence speed of the algorithm.Through the investigation of the CSC for 3D images with channel correlations,we find that existed methods can not consider the spatial-spectral intergrity of 3D images.This paper proposes a 3D convolutional dictionary learning algorithm based on the structure of 3D dictionary and 2D coefficients,which takes into account of spatial-spectral features.At the same time,the proposed preconditioned ADMM is used to solve the problem to avoid the calculation of complex inverse matrices.With the development of remote sensing technology,hyperspectral images(HSI)have been widely used.At present,the hyperspectral imaging technology based on the theory of compressive sensing is developed.The reconstruction from several compressive measurements is a very underdetermined inverse task.As an unsupervised feature extraction method,CSC can provide prior information for hyperspectral reconstruction.This paper proposes a high accuracy compressive chromo-tomography reconstruction algorithm based on convolutional sparse coding which combines the hyperspectral image reconstruction with the convolutional sparse coding algorithm,implements total variation regularization and finally improves the quality of the reconstructed images.Based on the proposed efficient convolutional dictionary learning algorithm using Preconditioned ADMM,the efficient 3D convolutional dictionary learning algorithm and the high accuracy compressive chromo-tomography reconstruction algorithm,an image reconstruction system is designed and developed,this system has the functions of 2D dictionary learning,3D dictionary learning,the reconstruction of disparity map,2D image restoration and the hyperspectral image reconstruction.