Enhancing Traffic Visibility Using Self Attention Based Generative Adversarial Network Model

Highway traffic monitoring systems often encounter difficulties caused by poorer image quality due to adverse weather conditions like rain, fog, insufficient lighting, and low visibility during nighttime. This paper introduces a fresh strategy to tackle these challenges by employing an advanced image enhancement algorithm centered on an enhanced Generative Adversarial Network (GAN) incorporating multiscale feature fusion. Our methodology integrates attention mechanisms within the generator network to effectively diminish noise and elevate focus on intricate details within high-frequency regions, thereby addressing issues such as poor resolution and blurred imagery. Furthermore, we refine the discriminator utilizing an improved PatchGAN methodology that employs a localized discrimination strategy, enhancing its ability to distinguish between generated and authentic images. By facilitating ongoing interaction between the generator and discriminator, we attain a state of Nash equilibrium, thereby preserving the integrity and authenticity of the restored images. Our proposed model delivers impressive results, boasting peak signal-to-noise ratio (PSNR) values of 32.5, 29.16, and 28.7, coupled with structural similarity index (SSIM) scores of 0.89, 0.85, and 0.84 for night, fog, and rain scenarios respectively. These outcomes underscore the efficacy of our approach in enhancing highway traffic imagery under adverse weather conditions, holding promising prospects for real-world traffic monitoring systems