Image Super-resolution Reconstruction Based On Convolutional Neural Network

Image super-resolution reconstruction is one of the most important research directions in the field of computer vision.It can effectively improve the visual effect of the image and can be used as a preprocessing step for other image algorithms.It is widely used in the field of public safety,medicine,and remote sensing imaging.In recent years,with the introduction and development of convolutional neural network methods,image super-resolution reconstruction has become the focus of researchers.Judging from the current research results,most researchers tend to study neural network model structure,achieve the purpose of improving the speed of the model and improving the effect of image super-resolution reconstruction through modifying the structure of the convolutional neural network.In this dissertation,starting from the convolutional neural network method,with the image super-resolution reconstruction task as the background,and based on the gradient magnitude similarity deviation（GMSD）function in image quality evaluation,a new loss function is proposed.So far,most of the convolutional neural network models in the field of image super-resolution reconstruction use thel ₂ loss function or the l₁ loss function to train the model,but using thel ₂ or l₁ loss function to train the model does not achieve the best results.This dissertation proposes a new mixed loss function based on GMSD.Using this loss function instead of the l₂ loss function and the l₁ loss function for convolutional neural network model training can improve the model's effect without changing the original model structure.