A new machine learning method cuts through the noise in laser optics inspection — telling real damage from convincing fakes with 96.36% accuracy.
High-power laser facilities photograph their final optics in real time, but those images routinely contain pseudo-damage sites that look nearly identical to genuine damage. Matching online detections to verified offline ground-truth records is the standard fix, but it has been unreliable: features between sites are sparse, local geometry warps, and decoy sites crowd the field. Existing models try to tune out those decoys through loss-function tricks alone, which leaves real-world performance shaky. The new confidence-feedback-weighted graph matching network takes a different approach — it only needs the centroid coordinates of each site, then iteratively estimates how matchable each node is, feeds that confidence score back as a weighting signal, and uses it to stop bad matches from polluting later steps.
The practical upside is more trustworthy damage accounting in facilities where a missed or misclassified site can mean the difference between a safe shot and a destroyed optic. Graph-based matching has matured quickly in computer vision, but applying it to a domain this narrow — centroid-only inputs, tiny datasets, high-stakes labels — is a meaningful test of how far the technique generalizes.
At 96.36% F1 on the team's own Complex-Scene dataset, the numbers look strong, though independent benchmarking on outside data would be the real proof of concept.