Colorization of Day-Night Aerial Infrared Images Using Online GIS Imagery

Humans find it difficult to recognize scenes in infrared images that lack color or contrast. A common approach is to convert infrared images into visible color images. Recent advances in deep learning have substantially enhanced image-colorization techniques. Simultaneously captured visible and thermal image pairs should be used for the training. However, acquiring such image pairs is challenging under low-light conditions, such as at night. To address this challenge, we propose a colorization method that uses Google Earth. Using GPS and camera orientation data, Google Earth can provide visible images with the same field of view as infrared images. We mitigated the differences in capture timing and angles between the infrared and visible images using a two-step approach: image colorization and image fusion. For image colorization, we employed a CycleGAN architecture with content loss to preserve the spatial structure while learning color mapping from visible reference images. For image fusion, we applied Multiscale Decomposition to extract and integrate fine structural details from infrared images. Experiments using SWIR and MWIR images demonstrated successful colorization in day and nighttime scenarios, confirming the effectiveness of our proposed method.