RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision
Researchers have developed a new method called RQUL-UIE to improve underwater image enhancement by addressing issues with unstable label quality in training data. The approach uses a diffusion-based, self-supervised learning strategy that leverages semantic perception embeddings from a pre-trained diffusion model to assess label quality. This allows the system to effectively utilize even low-quality labels while preventing them from degrading performance, and a Fourier-based network further refines high-frequency components for superior restoration. AI
IMPACT Improves image restoration techniques, potentially benefiting applications in marine research and underwater exploration.