Super-resolution Reconstruction Algorithm Based On Residual Structure And Dense Connection

High-resolution images usually have the advantages of being clear,stable,and rich in details.They are widely applied to education,military,and meteorology.However,limited by the existing optical image acquisition technology or affected by objective factors such as natural climate,the collected images often suffer from damage and noise.Image super-resolution reconstruction technology can regenerate low-resolution images to high-resolution images.The quality of reconstructed images obtained by traditional super-resolution reconstruction technology cannot always meet some specific needs,at the same time,the algorithm based on convolutional neural network can usually get excellent results by learning image structure.Thus,this paper researches super-resolution reconstruction technology based on convolutional neural network,aiming at further improving the quality of reconstructed images by ways of introducing residual connection structure,convolution block and dense connection structure.The main work includes:(1)This paper first briefly introduced the research background and significance of the super-resolution reconstruction algorithm.It discussed the existing super-resolution reconstruction technology,which can be divided into interpolation-based methods,reconstruction-based methods,and learning-based methods.The super-resolution reconstruction algorithm based on the convolutional neural network is to use the idea of deep learning for image reconstruction.Further,this paper discusses the basic principles of neural networks and the deficiencies in the SRCNN,such as oversized convolution kernel and too simple network model.For the deficiencies in the SRCNN,this paper proposes the super-resolution reconstruction algorithm based on the residual connection structure and the super-resolution reconstruction network based on dense connection structure.(2)The SRCNN mainly has the disadvantages that the convolution kernel is oversized and the network model is too simple.In this paper,a super-resolution reconstruction network.It stitches feature maps of multiple dimensions is designed by introducing a residual connection structure,which makes the neural network betters to extract structural features of low-resolution images.(3)Dense connection is another kind of cross-layer connection structure,and can maximize the use of the network's output feature map.Therefore,this paper designs a super-resolution reconstruction network based on dense connection structure between 3 sets of convolution blocks,so that the network model can better learn the mapping relationship between images structure.(4)In order to further verify the performance of the two super-resolution reconstruction algorithms,this paper will conduct the model training experiments and the comparison experiments of multiple super-resolution reconstruction algorithms.The training experiments is based on the computer and uses the Adam algorithm as the model training optimization algorithm,and the network model is built on the Pytorch deep learning framework.The comparative experiments firstly sets the upsampling parameters to 2 and 3,secondly,the reconstruction results of multiple algorithms are evaluated under two systems of subjective evaluation criteria and objective evaluation criteria,finally,the results prove that the two algorithms proposed in this paper have excellent image reconstruction performance.Overall,this paper proposes two algorithms based on convolutional neural network of residual connection structure,convolution block and dense connection structure,and the experiments prove these algorithms have excellent performance in reconstructing images.