Research On Compressed Image/video Sensing Algorithm Based On Deep Learning

The traditional images and videos compression coding framework is based on ShannonNyquist sampling theorem.First,the original signal is sampled at a frequency higher than twice the bandwidth of it.Then,the corresponding compression method is utilized to remove redundant information in the images and videos.This coding scheme inevitably brings a lot of waste in computing resources,and requires higher sensor resources during the sampling,which cannot be applied to some special scenarios,such as magnetic resonance imaging,wireless video surveillance,and so on.Compressed sensing theory breaks through the limitations of the Shannon-Nyquist sampling theorem in the traditional sampling framework,greatly decreasing the frequency requirements of the sampled signal.Besides,this theory completes compression at the same time during the sampling process,reducing the computational burden at the encoder.However,in traditional compressed sensing,the reconstruction end often uses complex iterative optimization algorithms for signal recovery.Such algorithms require massive running time,which hinders the practical application of compressed sensing.The recent proposal of deep learning based compressed sensing neural networks successfully solves the high time complexity problem existing in traditional algorithms,while achieving outstanding reconstruction performance.However,the research about the deep learning methods applied on image compressive sensing has just begun.The network structure is immature,and the reconstruction performance needs to be improved.Besides,the related research on video compressive sensing are even less.This paper focuses on the image and video compressive sensing methods based on deep learning.The main work includes the following three parts:1.A multi-stage image compressed sensing neural network(MSRes ICS)based on residual learning is proposed.On the one hand,the neural network successfully breaks through the traditional block compressed sensing framework during the sampling,and solved the block effect problem caused by it;on the other hand,it proposed a multi-stage image reconstruction neural network structure during reconstruction to improve the quality of restored images.Experimental results show that compared with other algorithms,MSRes ICS gains better reconstruction performance under various experimental conditions,and the average PSNR results are improved by 1-2d B.2.Based on the traditional SPL iterative optimization algorithm,a new image compression sensing neural network(SPLNet)is proposed.SPLNet effectively integrates traditional SPL algorithms into neural networks,which achieves theoretical interpretability and excellent reconstruction performance.Simulation results show that compared with state of the art reconstruction algorithm,SPLNet attains the best performance,while the average PSNR results are improved by 2-4d B.3.Based on SPLNet,a neural network(VCSNet)for video compression sensing reconstruction is proposed.In VCSNet,the time-domain information reconstruction subnetwork effectively extracts the correlation information in the video sequence and reconstructs a high-quality motion residual image,which acquires better reconstructed video signal.Compared with other traditional video compression sensing algorithms,VCSNet greatly reduces the runtime complexity of the reconstruction algorithms while maintaining good reconstruction performance.Compared with the existing video compressive algorithm based on deep learning(CSVideo Net),the reconstruction accuracy is effectively improved.