Unsupervised Translation from Shortwave Infrared Images to RGB Images: A Comparative Evaluation

Image-to-image translation has become an important technique in computer vision and focuses on preserving the effective content of images while enabling them to be translated from one domain to another. Among different spectral imaging formats, the Short Wave Infrared (SWIR) image format stands out with its ability to capture more information than the known color image format. However, the interpretability of SWIR images poses a significant challenge as they are not as easily understood by humans as color images. Translating SWIR images to Red-Green-Blue (RGB) images increases the interpretability of images by humans. It also allows models developed in the literature using color images to be used in translated color images. In this study, an original dataset is created where SWIR and color images can be translated into each other. Using this dataset, the performances of Cycle Consistent Adversarial Networks (CycleGAN), Contrastive Unpaired Translation (CUT) and Fast Constrastive Unpaired Translation (FastCUT) deep generative models, which are known to be successful in translating unpaired images, are compared. Models are evaluated according to both the quality of the images they generate and numerical metrics. As a result of the experiments, it is seen that the CUT model generates more successful results on the test data set in both types of evaluations.