Infrared-Augmented CycleGAN (IRACycle GAN) for Unpaired Night-to-Day Image Translation

This paper proposed Infrared-Augmented CycleGAN (IRACycle GAN), a novel extension of the traditional CycleGAN designed for unpaired Night-to-Day image translation. IRACycle GAN is a modified CycleGAN incorporating a 4-channel input architecture, simultaneously processing RGB and Infrared (IR) images. This dual-modality approach enables the model to capture visual features from the RGB images and critical thermal information, such as heat signatures and temperature variations in low light environments, by the IR images, significantly enhancing the realism and quality of generated Day RGB images. Unlike traditional CycleGAN, which relies solely on RGB data, IRACycle GAN's generator architecture processes RGB and IR inputs through separate pathways fused to create high-quality Day images. A key innovation is the modified cycle consistency loss, which ensures that thermal features from the IR input are preserved in the Day image, maintaining both visual fidelity and thermal context. Experimental results demonstrate IRACycle GAN’s superior performance, with an NIQE score of 4.5, a BRISQUE score of 39.32, and an FID score of 15.5. These results highlight IRACycle GAN's ability to generate more natural, realistic, and contextually accurate images by effectively integrating RGB and IR data. The model's enhanced performance makes it suitable for surveillance, autonomous vehicles, and thermal mapping applications, where visual and thermal information are critical for accurate interpretation and decision-making in challenging environments.