Dual discriminators generative adversarial networks for unsupervised infrared super-resolution

Abstract

The breakthrough of sensor resolution has always been focused on the computer vision area, which is an urgent need when infrared sensor technology progresses slowly. We propose a novel unsupervised infrared super-resolution (UISR) method based on the photoelectric characteristics of infrared images. The UISR, without high-resolution ground truth, establishes a fully unpaired generative adversarial framework to enrich the image details and improve the visual perception of human eyes. In the process of SR, UISR corrects the abnormal pixel values, blurred edges, and mosaic effect by maintaining the image style uniformity and mining the texture distribution characteristics in the image domain. Specifically, the dual discriminator module is designed to enhance both global structures and local textures. The style discriminator captures the natural structure characteristics of large image patches to control the authenticity. In contrast, the texture discriminator captures the texture characteristics of small image patches to improve the detail generation performance. Besides, the content constraint module is designed to maintain the basic content of the image. Experiments show that, compared to the supervised algorithm, the proposed UISR presents more real SR infrared images and is simpler to generalize in higher scales beyond sensor resolution.