Research And Application Of Image Super-resolution Reconstruction Algorithm Based On Symmetric Residual Connection

With the rapid development of digital image,image super-resolution reconstruction algorithm has been applied in medical diagnosis,remote sensing detection and video monitoring.The reconstruction method based on learning can obtain better reconstruction effect.However,as the depth of the network increases,the image super-resolution reconstruction method based on convolutional neural networks is prone to problems such as gradient disappearance and network degradation.To solve these problems,two image super-resolution reconstruction methods based on convolutional neural networks are proposed.The specific works are as follows:Aiming at the problems of gradient disappearance and network degradation in deep networks,an image super-resolution reconstruction method based on symmetric residual CNN is proposed.By setting a symmetric residual skip connection method in the network model,the network realizes the sharing of feature information within the residual block,which enhances the ability of extracting deep features of the image and reduces the amount of network parameters.The global feature fusion is realized by long-skip connection,which can effectively reduce the gradients disappearance and network degradation.The results of proposed reconstruction on Set5,Set14,and BSD100 perform superior to other methods in the comparison.The average PSNR and SSIM values are improved compared with those methods.Aiming at the characteristics of rich information and strong autocorrelation of remote sensing image,a super-resolution reconstruction algorithm based on symmetric local fusion block is proposed.This algorithm proposes a convolutional neural network structure of locally fused blocks,which improves the effect of reconstructing high-frequency information of remote sensing images.The network effectively alleviates the problem of insufficient high-frequency feature extraction of the deep network by setting local fusion in the block,and improves the reconstruction accuracy of the remote sensing image by deep network.The network uses the residual method to set skip connections between the blocks forming the symmetry,which improves global feature utilization and reduces network computational complexity.The results of proposed reconstruction on UC Merced and NWPU-RESISC45 are higher than other methods in the comparison.Both objective evaluation and subjective visual effects have achieved good reconstruction results.