Research On Image Compressed Reconstruction Based On Compressed Sensing And Deep Learning

The classic "first sampling-later compression" imaging mode causes a lot of waste of hardware resources,in the larger scale,higher dimensionality,wider dynamic range of data acquisition,processing,storage and transmission,lower power consumption,and imaging applications in more complex environments are facing huge challenges.Compressed imaging theory is based on the premise of sparsity or compressibility.It is expected to overcome some of the shortcomings of existing imaging theories.It can be applied to the traditional imaging field and is more suitable for complex occasions where traditional imaging theory is not applicable.However,there are several problems in the actual optical imaging system that must be considered.For example,the actual measurement matrix often depends on the physical properties of the measurement process.In many applications,it is required to obtain sufficient measurement values within a very limited time.And the performance and time limit of the reconstruction algorithm.The main work and innovations of this paper are as follows:(1)The reconstruction matrix used in the two-dimensional orthogonal matching pursuit algorithm is a measurement matrix,and the measurement matrix is an orthogonal matrix,which is difficult to realize physically and does not have optimal performance.The reconstruction matrix used in the two-dimensional orthogonal matching pursuit algorithm is measurement matrix,and the measurement matrix is an orthogonal matrix,which is difficult to obtain in an actual imaging system and does not have optimal performance.This paper introduces singular value decomposition into the reconstruction process of two-dimensional orthogonal matching pursuit algorithm,and proposes Optimized Two-Dimensional Orthogonal Matching Pursuit Algorithm via Singular Value Decomposition.We perform SVD on separable measurement matrices,optimized separable reconstruction matrices and optimized measurements are obtained,and then the original sparse signal can be reconstructed.The theoretical analysis and simulation experiments are provided to verify the effectiveness of the method in this paper.(2)The limitations of traditional Compressed Coded Aperture Imaging(CCAI)are thehuge demand for measurement matrix storage and the computational burden of algorithms to restore the original image.Therefore,CCAI is limited to relatively small-sized images.In view of the shortcomings that CCAI can only be used for small-scale images,this paper introduces Separable Compressed Sensing(SCS)into CCAI,and proposes Separable Compressed Coded Aperture Imaging(SCCAI)method.Then,we perform singular value decomposition on two separate measurement matrices,and optimize the measurement matrix and measurements,that is,the Separable Compressed Coded Aperture Imaging via Singular Value Decomposition(SCCAI-SVD)method.Finally,the superiority of the SCCAI-SVD method is verified by theoretical analysis and simulation experiments.(3)The iterative nature of traditional reconstruction algorithms leads to the limitation of reconstruction algorithms in practical applications,making it difficult to achieve real-time reconstruction.And the real image signal in nature is a compressible signal,which does not fully satisfy the sparseness condition,so the signal reconstruction quality is reduced in the actual imaging system.To solve this problem,deep neural network is introduced into the reconstruction process of compressed aperture imaging,and the deep learning architecture is trained by the data composed of images and measurements.The reconstruction process in the test phase is completed by non-iteration.Real-time reconstruction from low-resolution measurements.In addition,during the training phase,a pre-processing module is added to increase the matching degree between the forward operator and the backward operator to improve the reconstruction performance.The experimental results show that under the same experimental conditions,the method in this paper can achieve high-quality real-time reconstruction from low-resolution measurements.