Seeing through fibers: unsupervised image reconstruction in fiber bundle imaging systems

Fiber bundle imaging systems suffer from sampling artifacts such as honeycomb patterns due to their discrete and non-uniform fiber layout, fundamentally limiting image resolution. Conventional reconstruction methods rely on precise calibration of the fiber layout or learning from paired datasets, both of which have limited generalization across imaging setups and require sample-specific preparation. We present an unsupervised method for reconstructing high-resolution images using a burst of misaligned frames that does not require known fiber layout, paired training data, or per-sample calibration. Our approach jointly solves motion estimation and image reconstruction through test-time training. We model each burst frame as a deformed observation of a single canonical view, parameterizing the underlying motion with a coordinate-based network. A second coordinate-based network learns a joint super-resolved scene representation shared across aligned frames. Both networks are trained jointly end-to-end without paired ground truth or external supervision. Simulation and experimental results demonstrate that our method robustly removes fiber bundle artifacts and generalizes to various sample types. We also released a benchmark dataset for optical fiber bundle imaging to facilitate future research.