Research On Video Compressive Sensing Reconstruction Algorithm Based On Deep Network

The video compressed sensing system is based on the theory of compressed sensing,which projects multiple frames to two-dimensional compressed measurements in only one exposure process,and then realizes high-speed imaging.In order to recover the original video signal from the 2D compressed measurement signal,classical reconstruction algorithms are based on the sparse prior of the video for iterative optimization solution,but there are problems such as low reconstruction quality and time consuming.Deep learning has received wide attention for its excellent learning ability and has also been applied to the video compression-aware reconstruction problem,but the existing reconstruction algorithms based on depth methods lack effective representation for spatio-temporal features,and the reconstruction quality still needs to be improved.To this end,this dissertation focuses on deep learning models and algorithms for video compression-aware reconstruction,improves the spatio-temporal feature representation capability of the network,adopts a data-driven approach to learn the reflection projection process from the compressed measurement to the original signal,achieves fast reconstruction by simple feedforward calculation for new compressed measurement samples,and significantly reduces the reconstruction time complexity while ensuring the image reconstruction quaility.The details of the study are as follows:(1)A multi-scale fusion reconstruction network for video compression perception is proposed.Considering the multi-scale characteristics of video scene content,the network is developed from two dimensions: horizontal convolution depth and vertical resolution scale.The vertical dimension uses 3D convolution to extract video features at different scales,and the horizontal dimension uses the pseudo-3D convolution residual module to perform hierarchical depth abstraction of feature maps at the same resolution scale,and enhances the network's representation of video spatio-temporal features through cross-fusion of features at adjacent scales capability.The network model is guided by the joint constraints of reconstruction fidelity term,compressed measurement value fidelity term and total variance regularization term for learning parameters in the network.Experimental results show that the method can shorten the time of video reconstruction and improve the quality of reconstructed videos.(2)A residual integrated reconstruction network for video compression perception is proposed.Due to the spatio-temporal complexity of video scene contents,a single type of network is usually not able to characterize them effectively.For this reason,this thesis further constructs a residual integrated network model for video compression-aware reconstruction,which contains four sub-networks,namely,pseudo-3D convolutional codec sub-network A,serial bidirectional recursive sub-network B,branch residual sub-network C and main path residual sub-network D.The above four sub-networks are used to capture the complex scene appearance and object motion in the video.To effectively extract the spatio-temporal correlation between video frames,sub-network A uses a codec structure of pseudo-3D convolution,subnetwork B uses a serial bidirectional convolutional long-short memory network,and subnetworks C and D use a residual module for deep extraction of video features.These four subnetworks are connected by alternating residuals to form an overall network,which increases the number of information flow paths in the network,and then fully fuses the reconstruction results of each subnetwork to improve the reconstruction quality.The experimental results show that the network is able to achieve near real-time and high-quality reconstruction of the video.