The functionality of Autonomous Underwater Vehicles (AUVs) is contingent upon a range of sensors, encompassing acoustic, inertial, and visual modalities. Visual sensing emerges as particularly beneficial owing to its non-intrusive nature and the copious data it furnishes, most notably in shallow waters. Nevertheless, the integrity of underwater visual data is jeopardized by light refraction, absorption, suspended particulates, and color aberrations, which engender noise and distortion. In consequence, AUVs frequently grapple with challenges and manifest suboptimal performance in vision-based tasks. It follows that advancements in underwater image enhancement are paramount for marine engineering and aquatic robotics. This study introduces an original method that employs Generative Adversarial Networks (GANs) to enhance underwater optical scenes. By adapting CycleGAN, we have crafted specialized datasets for image restoration, thereby markedly elevating safety and operational efficiency in vision-guided AUVs. Both quantitative and qualitative evaluations corroborate the preeminence of our method. Experimental results reveal that our methodology outperforms extant techniques in PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) metrics, enhancing visual clarity and system reliability. Our discoveries highlight the potential of GAN-based approaches in mitigating underwater imaging obstacles, suggesting that further evolution and integration could revolutionize marine technology and aquatic robotics.