ZRLGANs: Zero-shot Robust Underwater Image Enhancement via Lightweight Compressed Generative Adversarial Networks

Abstract Underwater images often exhibit low quality due to factors like light scattering and impurities, challenging underwater detection and exploration. While deep learning advances, such as GANs, have improved image enhancement, they face limitations in lightweight applications due to high computational demands. We propose a zero-shot robust multi-teacher knowledge distillation-based compression approach, ZRLGANs, specifically designed for underwater image enhancement. ZRLGANs significantly reduces computational complexity by 13 times and parameter count by 18 times, as demonstrated on EUVP, UFO-120, and UIEB datasets, while maintaining real-time performance and minimal quality loss. This method effectively balances efficiency and quality for resource-limited deployments.