Research On Fast Image Restoration Method Based On Symmetric Expanded Convolutional Network

At the stage of image data acquisition,decompression,and transmission,the collected image data is degraded due to hardware and environment.Degraded images lose some important information,which reduces the accuracy of information processing.Image restoration is an ill-posed problem in mathematics,meaning there are infinite solutions.The reconstructed image can only be finitely close to the original image,so image restoration is a research difficulty in the field of image processing.This paper focuses on image super-resolution and image denoising.In recent years,due to the powerful feature extraction capabilities and extensive application scenarios of deep learning,image restoration methods based on convolutional neural networks have developed rapidly and become the current popular models.Existing image super-resolution and image denoising methods based on convolutional neural networks still have optimization space,and the speed and accuracy can further improved.For the above problems,this paper works as follows.(1)First,the proposed method uses residual learning as the main structure,which learn residual image between degraded image and original image.In order to improve the speed of the model,this paper upgrades dilated convolution,which achieves a larger receptive field with fewer convolution layers and avoids griding phenomenon of dilated convolution.(2)In addition,batch normalization combined with residual learning in image denoising can effectively fit Additive White Gaussian Noise.However,batch normalization costs more computation,and image super-resolution does not follow this rule.Therefore,this paper attempts to remove the batch normalization and replace it with symmetric skip connections.The experimental results show that the proposed network(SDNet)can effectively deal with image super-resolution and image denoising.(3)Based on SDNet,this paper proposes an enhanced version SDNets,which add a little calculation.According to the characteristics of image super-resolution and image denoising,SDNets enhances the two tasks separately.Experiments show that SDNets have higher PSNR than SDNet.This paper compares SDNet/SDNets with mainstream image restoration methods,including the comparison method of objective evaluation and subjective evaluation,and also makes a simple comparison in several important training sessions.In the general test set,the proposed method is superior to the mainstream image restoration methods in terms of accuracy,speed,and visual.The biggest advantage of SDNet/SDNets is that the calculation speed is very fast,the CPU speed is twice as fast as the similar method,and the GPU speed is also fast.