Light the Way: An Enhanced Generative Adversarial Network Framework for Night-to-Day Image Translation With Improved Quality

Driving at night introduces considerable challenges due to reduced visibility, making it essential to explore techniques that enhance road information for drivers. With this purview, the research presents a technique to address visibility constraints faced during night-time driving, by converting night-time road images to day-time images using a supervised Generative Adversarial Network (GAN) model (NtD-GAN). Since paired images are required to train supervised GAN models, the research first exploits a novel approach for generating paired night-day datasets, as it is practically infeasible to collect such image pairs in a natural setting, owing to dynamic traffic environments. An innovative generator network architecture is proposed for the NtD-GAN. Furthermore, a new approach was proposed for generating and loading initial weights to expedite the NtD-GAN training. This initial weight assignment resulted in faster convergence of the NtD-GAN with significant improvement in Inception Score (IS) by 17.3%, in Structural Similarity Index (SSIM) by 5.5%, and in Naturalness Image Quality Evaluator (NIQE) by 10.3%. Moreover, the perceptual loss is introduced to the training loss function of the NtD-GAN to increase the visual quality of the reconstructed images. The experimental results also demonstrated a 0.23% increment in IS, a 0.07% reduction in Fréchet Inception Distance (FID), a 2.2% increment in SSIM, and a 7% reduction in Blind Referenceless Image Spatial Quality Evaluator (BRISQUE) compared to the NtD-GAN trained without perceptual loss. The comparison analysis with the benchmark models has demonstrated a significant improvement. For instance, in comparison to N2D-GAN, NtD-GAN has demonstrated a reduction in FID by 14.6%, an improvement in SSIM by 3.4%, an improvement in Peak Signal-to-Noise Ratio (PSNR) by 1.39 dB and a reduction in BRISQUE by 0.8%. The implementation of the NtD-GAN model is available at https://github.com/isurushanaka/paired-N2D.