Learning to reconstruct from saturated data: audio declipping and high-dynamic range imaging
#self-supervised learning #audio declipping #high-dynamic range imaging #signal reconstruction #machine learning #inverse problems #clipped measurements
📌 Key Takeaways
- Researchers extended self-supervised learning to non-linear problems of recovering audio and images from clipped measurements
- The method assumes signal distribution invariance to amplitude changes
- Experiments show the approach is almost as effective as fully supervised methods
- The technique works with clipped measurements alone, eliminating need for ground truth
📖 Full Retelling
Researchers Victor Sechaud, Laurent Jacques, Patrice Abry, and Julián Tachella have developed a novel approach to reconstruct audio and images from clipped measurements, addressing a critical limitation in current machine learning applications, in a paper submitted to arXiv on February 25, 2026. The research tackles the fundamental challenge of deploying learning-based methods for solving inverse problems in real-world scenarios where ground truth references for training are unavailable. While recent self-supervised learning strategies offer promising alternatives by eliminating the need for ground truth, most existing approaches are constrained to linear inverse problems only. The researchers' breakthrough extends these capabilities to the non-linear domain of recovering audio and images from clipped measurements, assuming that signal distributions remain approximately invariant to amplitude changes. Their work establishes sufficient conditions for learning reconstruction from saturated signals alone and introduces a self-supervised loss function that can effectively train reconstruction networks. Experimental results demonstrate that the proposed approach achieves nearly equivalent performance to fully supervised methods despite relying exclusively on clipped measurements for training, opening new possibilities for applications in audio restoration and high-dynamic range imaging.
🏷️ Themes
Machine Learning, Signal Processing, Self-Supervised Learning
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
--> Electrical Engineering and Systems Science > Image and Video Processing arXiv:2602.22279 [Submitted on 25 Feb 2026] Title: Learning to reconstruct from saturated data: audio declipping and high-dynamic range imaging Authors: Victor Sechaud , Laurent Jacques , Patrice Abry , Julián Tachella View a PDF of the paper titled Learning to reconstruct from saturated data: audio declipping and high-dynamic range imaging, by Victor Sechaud and 3 other authors View PDF HTML Abstract: Learning based methods are now ubiquitous for solving inverse problems, but their deployment in real-world applications is often hindered by the lack of ground truth references for training. Recent self-supervised learning strategies offer a promising alternative, avoiding the need for ground truth. However, most existing methods are limited to linear inverse problems. This work extends self-supervised learning to the non-linear problem of recovering audio and images from clipped measurements, by assuming that the signal distribution is approximately invariant to changes in amplitude. We provide sufficient conditions for learning to reconstruct from saturated signals alone and a self-supervised loss that can be used to train reconstruction networks. Experiments on both audio and image data show that the proposed approach is almost as effective as fully supervised approaches, despite relying solely on clipped measurements for training. Subjects: Image and Video Processing (eess.IV) ; Artificial Intelligence (cs.AI cs.SD) Cite as: arXiv:2602.22279 [eess.IV] (or arXiv:2602.22279v1 [eess.IV] for this version) https://doi.org/10.48550/arXiv.2602.22279 Focus to learn more arXiv-issued DOI via DataCite Submission history From: Victor Sechaud [ view email ] [v1] Wed, 25 Feb 2026 10:37:14 UTC (17,431 KB) Full-text links: Access Paper: View a PDF of the paper titled Learning to reconstruct from saturated data: audio declipping and high-dynamic range imaging, by Victor Sechaud and 3 other authors View PD...
Read full article at source