Colorization of thermal infrared images to low-light mode using cGAN

The focus on this research is to automatically generate color representations in red, green, blue (RGB) channels from thermal infrared (TIR) images as input that only contain grayscale intensity information and are acquired under low-light mode. This process aims to improve the visual interpretability of TIR images by producing outputs similar to those obtained in the visible spectrum especially in low-light modes. This study employs a paired dataset comprising TIR and corresponding visible-spectrum images, both captured under nighttime conditions. The proposed method utilizes a Conditional Generative Adversarial Network (cGAN), which includes two main networks: a generator and a discriminator. Both networks process adversarially, where the generator produces TIR image that resembles the ground truth image, while the discriminator distinguishes between the image generated and the ground truth image. The generator adopts a U-Net architecture to preserve the spatial details of the inputted image through the use of skip connections to combine the spatial information from the encoder with the decoder features. Meanwhile, the discriminator employs a PatchGAN architecture to evaluates the realism of the generated color images at the level of local image patches. Model performance is evaluated using three metrics to obtain the best results on the testing data. The evaluation metrics are MSE, PSNR, and SSIM which reached values of 0.003, 25.97 dB, and 0.78.