Zero-Shot Learning for Haze Removal Using Fusion of Near-Infrared and Color Images

Outdoor images are often degraded by haze. In this study, we propose a new method to remove haze by fusing near-infrared (NIR) images and color images using zero-shot learning. Unlike most learning-based methods that require a large amount of training data, our method employs zero-shot learning, which learns using the input image only. The proposed method removes haze from both NIR and color images. In edge fusion, a weight map is generated and added to the NIR image, which is then multiplied by the lightness image of the haze-removed color image. This process reduces the differences in characteristics between the NIR and lightness images, achieving a natural fusion of edges. Our qualitative and quantitative evaluations using 59 pairs of NIR and color images demonstrated that our method can restore color and edges more naturally than conventional methods.