Infrared image super-resolution using auxiliary convolutional neural network and visible image under low-light conditions

Convolutional neural networks (CNN) have been successfully applied to visible image super-resolution (SR) methods. In this study, we propose a CNN-based SR algorithm for up-scaling near-infrared (NIR) images under low-light conditions, using corresponding visible images. Our algorithm first extracts high-frequency (HF) components from the up-scaled low-resolution (LR) NIR image and its corresponding high-resolution (HR) visible image, and then takes them as multiple inputs of the CNN. Next, the CNN outputs the HR HF component of the input NIR image. Finally, an HR NIR image is synthesized by adding the HR HF component to the up-scaled LR NIR image. The simulation results show that the proposed algorithm outperforms the state-of-the-art methods, in terms of both qualitative and quantitative aspects.