Illuminating the Roads: Night-to-Day Image Translation for Improved Visibility at Night

Image enhancement and night-to-day image translation are the most prominent approaches for improving visibility at night. Yet, the capability of improving visibility using image enhancement techniques, such as gamma correction, is limited. On the other hand, translating night-time images into day-time images has more potential and modern deep learning techniques such as Variational Auto Encoders (VAEs) and Generative Adversarial Networks (GAN) can be utilized for this purpose. This study focuses on exploring the possibility of achieving night-to-day image translation with a supervised GAN. To enable the training of a supervised GAN model, the availability of a pixel-to-pixel paired night-day image dataset is crucial. Hence, a pixel-to-pixel paired night-day image dataset was generated by pairing through synthesis using an existing pre-trained model. Subsequently, the proposed supervised GAN model was trained on the generated dataset. To facilitate a comprehensive comparison, a well-established pre-trained night-to-day image translation model from ToDayGAN was selected as the benchmark. According to the comparison analysis, the supervised GAN model provided 16.32 dB of average Peak-Signal to Noise Ratio (PSNR) improvement, 0.513 of average Structural Similarity Index (SSIM) improvement, 0.061 of average Inception Score (IS) improvement and 67.33 of average Fréchet Inception Distance (FID) reduction compared to the ToDayGAN night-to-day image translation model.