Compressed Sensing Reconstruction Of Image Sequences Based On Unsupervised Network Learning

Compressed Sensing(CS)technology fuses sampling and compression processes to reconstruct the original signal from measurements much lower than Nyquist sampling.It has been successfully applied to image sequence imaging systems,such as Coded Aperture Spectral Imaging(CASSI),which greatly improves imaging efficiency.Reconstruction algorithm is a topic that has been focused on in the field of CS research.At present,CS reconstruction algorithms based on deep learning have received widespread attention due to their high reconstruction quality,but these algorithms are usually supervised learning and require network pre-training on a large number of data sets.At the same time,for different measurement rates,measurement matrices and imaging models,the network needs to be retrained again,which is not conducive to the deployment of the reconstructed network in practical applications.Therefore,this dissertation focuses on the unsupervised network learning of CS reconstruction,and discusses the problems of CS reconstruction of image sequences such as hyperspectral images,natural images and videos.The specific research results include:(1)An unsupervised spatial-spectral network learning algorithm for hyperspectral compressed snapshot reconstruction is proposed.Based on the CASSI method,a generative network based on compressed snapshot measurement is constructed to represent the parametric mapping of hyperspectral reconstruction.Considering the spatial-spectral correlation of hyperspectral images,the spatial-spectral joint attention module is designed in the conditional generative network,and the effective expression of the spatial-spectral information is realized through the multi-scale three-dimensional attention mechanism.In the learning of the objective function,the network parameters are optimized,so that the hyperspectral image generated by the network matches the given compressed snapshot measurement,and high-quality reconstruction can be achieved without the need for the pre-training process of the network.(2)A CS reconstruction network based on the non-local regularization constraint is proposed.In order to enhance the robustness to noise in the unsupervised reconstruction process,the non-local regular prior of the image sequence and the deep network prior are further combined to learn the reconstruction.A half-quadratic splitting algorithm is designed to solve the compound constraint problem,and the original problem is decoupled into two sub-problems to be solved alternately by introducing auxiliary variables.The first sub-problem only involves the learning of network parameters,and the second sub-problem corresponds to the non-local regularization learning of the image sequence.Through a large number of experiments on image sequences such as natural images,videos and hyperspectral images,it is verified that this algorithm can maintain robustness to noise while improving the reconstruction accuracy.