Multi-channel Image Super-resolution Method Based On Deep Convolutional Neural Network

Multi-channel image super-resolution is the basis of image analysis and semantic understanding.Due to the influence of natural conditions and the interference of imaging device devices,multi-channel images obtained by imaging often have reduced resolution and fuzzy noise.Super-resolution reconstruction is the use of algorithms to reconstruct these low-resolution images into high-resolution images.The multi-channel image based on convolutional neural network simulates the structure of human brain neuron synapses,and uses a large amount of training data to learn the mapping weight matrix of low-resolution images and high-resolution images.At the same time,since the multi-channel image has time or spectral dimension compared to the single image,the adjacent image is applied to the convolutional neural network training to supplement the information of the intermediate image by using spatiotemporal correlation or spatial correlation,to obtain more Good super resolution.The thesis focuses on the multi-channel image super-resolution method based on deep convolutional neural network,and achieves the following results:(1)The VSRnet model for video super-resolution using adjacent inter-frame redundancy information is studied,and the model is generalized in three aspects to obtain the DeepVSRnet model.Firstly,the number of model input images is extended from 3 frames to K frames,and then the number of feature extraction layers before concatenation of images in the convolutional neural network is extended from 1 layer to N layer,and finally the images in the convolutional neural network are cascaded.The number of reconstruction layers is extended from layer 1 to layer M.The effects of network layer changes,input frame number changes and loss function optimization methods on model performance are discussed through experiments.(2)Combining the good results of convolutional neural network in optical flow estimation,the FlowNet2 network is applied to motion compensation,and the end-to-end video super-resolution model DeepVSRnet2 based on convolutional neural network is constructed.(3)In order to make better use of the supplementary information between adjacent frames,a loss function based on spatiotemporal correlation is constructed.Firstly,the motion trajectories of the super-resolution video image and the real video image are calculated separately,and then the motion trajectory consistency is used as a new loss function.Finally,the constructed loss function is applied to the training of the video super-resolution reconstruction algorithm DeepVSRnet2 model.This can get better video super resolution.(4)The DeepVSRnet model in video super-resolution is extended to hyperspectral image super-resolution.Because the hyperspectral image and the video image have similarities and differences,the DeepVSRnet model is modified to meet the hyper-resolution image super-resolution requirements.A hyperspectral image based on an optical spectrum joint and deep convolutional neural network is proposed.Resolution reconstruction model.