Learning to reconstruct from saturated data: audio declipping and high-dynamic range imaging

Researchers have developed a self-supervised learning method that can reconstruct audio and images from clipped/saturated measurements without requiring ground truth training data. The approach extends self-supervised learning to non-linear inverse problems and performs nearly as well as fully supervised methods while using only clipped measurements for training.

• →Self-supervised learning has been extended to handle non-linear inverse problems in signal reconstruction.
• →The method can recover audio and images from saturated data without needing ground truth references for training.
• →The approach assumes signal distribution remains approximately invariant to amplitude changes.
• →Experimental results show performance nearly matching fully supervised approaches despite using only clipped measurements.
• →The research addresses a key limitation in deploying learning-based methods for real-world inverse problems.