Improved Double Regression Nonlinear Image Super Resolution Model

Abstract The existing super resolution reconstruction methods are mainly divided into traditional and deep learning reconstruction. main problem faced by algorithms, such as image enlargement space transformation, is how to establish the mapping relationship between input target image, express pixel value of through relationship. As a prominent problem, difficulty lies in fact that there no realizable matrix one - many relationships. Based on U-Net network framework, this paper improves jump-connected modules. By using combination convolutional layer, activation layer residual channel block, overall module operation efficiency increased 2.4%, PNSR 0.49db, running speed 0.3ms average when processing single compared with other classical models.