Nonlinear image interpolation via deep neural network
Wentian Zhou, Xin Li, Daryl S. Reynolds · 2017
Model-based approaches toward image interpolation attempt to predict unknown pixels at the high-resolution (HR) from a given low-resolution (LR) image. In the past decade, various adaptive image interpolation methods have been develop aiming at better recovering important image structures such as edges and textures. However, those methods are all based on an implicit assumption about the linear relationship between LR and HR pixels partially due to the difficulty with modeling nonlinear relationship. In this paper, we propose to take an explicit learning-based approach toward modeling the nonlinear relationship between LR an HR pixels. A six-layer convolutional neural network with rectified linear units (ReLU) is presented and trained to learn the targeted nonlinear mapping from training data for image interpolation. Our experimental results have shown the proposed learning-based approach is often capable of achieving superior performance both subjectively and objectively to existing model-based methods.