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

Area of Science:

• Optics and Imaging
• Computer Vision
• Image Processing

Background:

• Fiber bundle imaging systems are limited by sampling artifacts like honeycomb patterns, reducing image resolution.
• Conventional reconstruction methods require precise calibration or paired datasets, limiting their generalizability and requiring sample-specific preparation.

Purpose of the Study:

• To develop an unsupervised method for reconstructing high-resolution images from fiber bundle imaging systems.
• To overcome limitations of conventional methods by eliminating the need for known fiber layout, paired data, or per-sample calibration.

Main Methods:

• Utilized a burst of misaligned frames for reconstruction.
• Employed test-time training to jointly solve motion estimation and image reconstruction.
• Modeled frames as deformed observations of a canonical view using coordinate-based networks for motion parameterization and scene representation.

Main Results:

• Successfully removed fiber bundle artifacts, such as honeycomb patterns.
• Demonstrated robust generalization across various sample types in simulations and experiments.
• Achieved high-resolution image reconstruction without ground truth or external supervision.

Conclusions:

• The unsupervised method offers a generalizable solution for high-resolution imaging with fiber bundles.
• The approach effectively addresses sampling artifacts without complex calibration or data requirements.
• A new benchmark dataset for optical fiber bundle imaging was released to advance research.