Research On Single Image Super Resolution Reconstruction Technology Based On Deep Learning

With the advent of the digital information age,images have become indispensable in all walks of life as the main information carrier.However,due to the hardware limitations of the imaging system and the influence of external conditions,people usually obtain low-resolution images,and high-resolution images have high requirements in medical images,digital television,and satellite imaging.Single-image super-resolution(SISR)is a well-defined computer vision problem.It aims to reconstruct a high resolution image from a single low resolution image.Recent studies have shown that image super-resolution methods based on deep learning can significantly improve the performance.In particular,residual networks and dense structures has made it possible to train deep deconvolution networks.Both of these are skip connection.The residual structure conveys the residuals.The dense structure conveys the feature map of the network.They can improve the network gradient circulation and information exchange.Therefore,the SISR method based on deep residual or deep dense structure has achieved great breakthroughs in using extremely deep network structures.In this paper,we use the higher order recurrent neural network theoretically relate the residual structure to the dense structure,and proved that these two structures have their own advantages and disadvantages.The residual structure focuses on the reuse of features,while the dense structure focuses on the exploration of new features.In order to combine the advantages of these two structures,strengthen the feature extraction function of deep neural network,we designs a dual patn network model structure combining residual and dense channels with modular structure,the feature maps are split into two paths,one path is propagated in the form of residual,and another path is propagated by dense skip connections,At the same time,the network introduces a deconvolution layer for feature amplification,effectively reducing the computational load of the network by about 4 times.Based on this,for the SISR task,this paper designed a 60-layer deep deep convolutional neural network.At the same time,in order to optimize the network training,we makes the optimal exploration choices in dataset,loss function,activ ation function,optimization algorithm and network initialization algorithm.We uses the two commonly used image quality metrics,PSNR and SSIM,compared with other SISR methods in recent years on four international common standard datasets.And under the same experimental conditions,three advanced network structure models were reproduced for image detail comparison.The experimental results show that the reconstructed image obtained by the network structure proposed in this paper is closer to the original image,and the reconstruction performance of the texture is clearer.On the Set5 test set,PSNR is improved by about 0.1dB~1.6dB compared to other deep learning methods,and 3.7dB is improved compared to the traditional non-depth learning method.Therefore,the proposed dual-path network model can structure the advantages of the two channels,enhance the network feature extraction capability,and achieve the best reconstruction performance.