Research On Up-sampling Layer Of Image Super-Resolution Network

With the rapid development of computer science technology and image processing technology,using image Super-Resolution algorithm to improve image resolution has important significance and application value.Image super-resolution is an algorithm that reconstructs from one or a group of low resolution images to corresponding high resolution images.The mainstream image SR algorithms are SR neural networks,including FSRCNN,Lap SRN,and ESPCN.The FSRCNN has less parameter and effect and the Lap SRN has better parameters and effect,with the ESPCN between them.To improve the HR image effect of the ESPCN model,we optimizes its upsampling layer by replacing the upsampling method and using the cascaded upsampling layer,and experiment with Tensorflow.It shows that the scheme using cascaded deconvolution layers as the upsampling layer can achieve better reconstruction results.The PSNR of the reconstructed image is 27.9db,which is 3.8db higher than that of the original low-resolution image,2.1db higher than Bicubic interpolation,and 0.27 db higher than the original ESPCN model.