Research On Compressed Sensing Algorithm Based On Convolutional Neural Network

Compressive sensing(CS)is an sampling method which recovers signals from random measurements.It is widely used in image processing,geophysics,medical imaging,computer science,signal processing and mathematics.This paper mainly studies the reconstruction algorithm based on deep network.I proposes two algorithm models and evaluates the reconstruction performance from the aspects of the sparsity dependence condition,the dimension of measurement value,running time and anti-noise ability.The main content and innovation points of this paper can be summarized as follows:1.Three sparse dictionary models of Gaussian pulse functions and ksvd learning dictionary are proposed for soil UWB signals.In the experiments,the first-order Gaussian pulse dictionary is more stable and effective than other dictionaries.2.A distributed compression sensing model based on long-short term memory networks(LSTM)network is designed.In simulation,the joint sparse prior condition is studied.It is proved that the compression sensing for deep learning reconstruct original signals from measurement without relying on sparse prior information,which provides the research basis for future models.3.According to the principle of orthogonal matching pursuit(OMP)algorithm,a set of compression perception reconstruction model(OMP-IRCNN)based on residual convolution neural network is designed.Then,an effective method and training method for generating training data set are propose d.The reconstruction experiment shows that the model is superior than other algorithms in the aspects of reconstruction error,compression rate,test time.And the model does not depend on the sparse prior condition.4.This paper analyzes the advantages of transfer learning in the field of compressed sensing.Then I establishes a convolution-based transfer compressed sensing algorithm(CTCS),using fine-tune method and minimize Multi-Kernel maximum mean discrepancies loss function in high-dimensional convolution layer.And the model transfers the source domain model to the target domain.In the experiment,CTCS modle realizes high-performance reconstruction in the target signals.Then,the advantages of CTCS algorithm in noise robustness,reconstructionerror,dimension of compression measurement and sparse prior condition restriction are analyzed and verified.