Image Compressive Sensing Reconstruction Based On Cascaded Network

Compressive sensing theory breaks through the constraints of the traditional Nyquist sampling theorem,and proves that sparse signals can be accurately reconstructed with the measured values of their reduced-dimensional sampling.Compressive imaging is an important application direction of the theory of compressive sensing,which has broad application prospects in the fields of magnetic resonance imaging,remote sensing imaging and radar imaging.The effective reconstruction of compressed sensing images is one of the core concerns of compressive imaging applications.At present,the deep learning based method provides a good potential for the fast and high quality reconstruction of compressive sensing images,and the related research has aroused great interest among researchers in this field.This paper is aimed at the real-time application of compression imaging,focusing on the research of image compression sensing reconstruction method which based on deep network,especially the method that based on convolutional neural network and deep model unfolding,and achieving some meaningful results in improving the quality of reconstructed image and effectively reducing the difficulty of deep network training.The work content is as follows:1)Research on the image compressive sensing reconstruction network with cascaded model unfolding and residual learning.Usually the depth of deep network is increased,the system performance will be improved to some extent.However,the deepening of the network depth will make the network training more difficult.On the basis of researching the deep model unfolding network,propose to reduce the number of layers of the original deep model unfolding network,and cascade an independent residual denoising network at the back end,to both improve the image reconstruction quality and effectively reduces the difficulty of training.Firstly,the front end deep model unfolding network which inherits the advantage of character well interpretability of the model optimization method,solves the initial reconstructed image from the compressed measurement value.The back end residual denoising network takes the initial reconstructed image output from the front end as input,and performs deep denoising to obtain higher quality reconstruction results.The training of the two level cascade network is completed independently,and the training process is simple and easy to implement.2)Research on the image compressive sensing reconstruction network which based on adaptive sampling and multi-scale residual learning.The size of the convolution kernel of the convolutional neural network determines the receptive field of feature extraction.Therefore,to improve the reconstruction effect,the reconstruction network which use small-scale convolution kernels,usually by adding the network depth to extract more relevant information.For the conventional compressed sampling which use the fixed coefficient measurement matrix,since the statistical characteristics of the sampled image cannot be adaptive,the sampling efficiency and the reconstruction effect are limited.Research a compressive sensing reconstruction method which based on adaptive sampling and multi-scale residual learning.Firstly,adaptive sampling reconstruction is used to obtain the initial reconstruction result of the image.Adaptive sampling improves the image sampling efficiency.Then,multi-scale feature extraction and deep reconstruction are performed on the initial reconstructed image through the multi-scale residual learning network.By combining different size convolution kernels and residual networks,not only reduce the number of network layers,but also reduce the difficulty of network training.At the same time,effectively expands the receptive field of feature extraction,and improves the image reconstruction quality.A large number of experimental results are given,to verify the effectiveness of the proposed method in improving the quality of reconstructed images and reducing the difficulty of network training.