Traditional multi-dimensional signal compressed sensing methods will generate a lot of redundant information,resulting in huge consumption of storage space,slow reconstruction speed will also consume a lot of time and resources,especially at low sampling rates,it is difficult to effectively reconstruct the original signal.The effective sampling and rapid reconstruction of signals are the key to solving the challenges of large-scale,high-dimensional,wide-dynamic and low-consumption information transmission.As an efficient and intelligent data and signal processing method,deep learning can automatically and effectively extract signal features and play an important role in easily and quickly reconstructing the original signal.To address the problems of long time and low quality in reconstruction of traditional multidimensional signal compression perception,this paper applies deep learning technology to the field of multi-dimensional signal compressed sensing,learning the internal correlation of multi-dimensional signals,and can use a small amount of measurement values to quickly and efficiently reconstruct the original multi-dimensional signal.The main work and research of this paper are as follows:(1)A multi-channel compressed sensing and optimization method based on singular value decomposition is proposed.Since the traditional compressed sensing method requires the orthogonality of the measurement matrix while ensuring the orthogonality of the reconstruction matrix,it greatly increases the complexity of the imaging system design.In the process of compressed sensing,singular value decomposition is used to preprocess the measurement matrix,and the reconstruction matrix and the measurement matrix are designed separately.The optimized reconstruction matrix rows are orthogonal to each other,which eliminates the correlation between the measurements and reduces the system design complexity while improving the signal reconstruction performance.The relevant theoretical analysis and comparative experiments are also provided to verify the superiority of the method.(2)A real-time reconstruction method of color compressed imaging and depth unfolding is proposed.To address the problems of long time and low accuracy of reconstruction in traditional color image compression reconstruction methods,this paper combines compressed sensing and deep learning,and made the following improvements.First,we extract the R,G and B components of color images and train them separately using the depth unified unfolding network.And applying the idea of migration learning,on the basis of completing the training of R channel images,we further complete the training of G and B channel images,which not only improves the reconstruction performance of color image,but also greatly reduces the network training time and saves a lot of resources and time.Then,an end-to-end unified training network is proposed,and the three-channel images are trained in a unified manner,avoiding the waste of resources caused by multiple network training.In addition,this paper adds a pre-processing layer to the network to obtain optimized measurements and measurement matrices,which reduces the complexity of the front-end data acquisition system design.A large number of experimental results show that this method has better performance than existing methods.(3)A deep video compressed reconstruction method based on non-iterative unfolding network is proposed.As traditional video reconstruction methods require sparse prior knowledge and a large number of iterations,their large computational complexity and long reconstruction time cannot meet the requirements of real-time high-quality reconstruction,which greatly limits their application scope.In this paper,the non-iterative unfolding network is further applied to the compressed reconstruction of video,and the redundancy of time dimension existing between video frames enables the model to better learn the correlation between training samples,and the video frames can be quickly reconstructed by low-dimensional measurements.In addition,the training of the depth network is completed by obtaining optimized video frames and measurements from the singular value decomposition preprocessing layer,which effectively improves the performance of video reconstruction.Experimental results illustrate that the method in this paper can complete high-quality real-time reconstruction of video from a small number of measurements. |