Snapshot Compressive Imaging Based On Deep Unfolding

Snapshot Compressive Imaging(SCI)aims to capture 3-dimensional data using a2-dimensional sensor.The basic principle is that a mask is used to modulate the highspeed scene during exposure in order to obtain compressed measurement frames.In this way,each measurement frame is made to contain all the information of the encoded frame,and then the original data is recovered using a reconstruction algorithm.Due to the advantages in data storage,transmission,SCI technology has a wide range of application prospects in the field of high-speed photography.Currently,algorithms for solving SCI problems are divided into four main categories: traditional optimization algorithms,classical deep learning algorithms,plug-and-play algorithms,and deep unfolding algorithms.At present,the existing reconstruction methods have disadvantages such as weak mask adaptation capability,long reconstruction time and poor interpretability.In this thesis,I will start from the unfolding strategy of the deep unfolding algorithm to improve the comprehensive performance and mask adaptation ability of the model.In SCI systems,model-based and learning-based methods have been used to reconstruct video frames from compressed samples,but most of these methods cannot achieve a good balance between accuracy and speed.For SCI reconstruction,generalized alternate projection(GAP)is one of the most effective algorithms.To enhance the SCI reconstruction performance,a deep unfolding network called GAPMSF-Net is proposed in this thesis.First,motivated by the fact the accelerated computation process of the original GAP algorithm can significantly improve the convergence speed of the algorithm,a deep accelerated network is designed based on its computational principle,which can learn a rule to adaptively adjust the linear manifold in a data-driven manner based on the reconstruction results of the previous stage.In addition,to enhance the information interaction between different reconstruction stages in the deep deconvolution algorithm,a multi-stage deep prior fusion is designed in this thesis.At each reconstruction stage,the deep prior fuses and interacts with the hidden states that take all the previous information.The superiority of the proposed method is demonstrated on simulated and real datasets in terms of accuracy and computational complexity.Currently,end-to-end network models require mask information to train,but the use of different masks across SCI systems means that changing the masks requires retraining models,leading to longer deployment times.Taking inspiration from traditional optimization algorithms and considering that tensor representation can store more structural information than vectors and matrices in multiple frames of data,this thesis extends the approximate message passing algorithm to tensor form to achieve SCI reconstruction tasks,named Tensor AMP-Net.The model unfolds the algorithm through a denoising perspective and uses a two-stage denoising unrolling strategy.The model has a simple structure and a small number of parameters for the network,which allows the model to inherit the properties of traditional optimization algorithms and enhance the robustness of the mask.In addition,considering the data structure characteristics between and within video frames,a deep prior module is implemented using a spatiotemporal convolutional network structure to achieve dual denoising functions from both local and global dimensions.Experimental results show that the model has only a slight degradation in reconstruction on the test mask compared with the training mask,and has a significant advantage in the speed,memory consumption and computation of the reconstruction.