PRISM-UGF: weakly paired LIBS-image fusion for local burn inspection of stainless steel

Strict point-to-point correspondence between surface images and LIBS spectra is difficult to establish in local burn inspection. We propose position-aware regular image sampling with uncertainty-gated fusion (PRISM-UGF), a weakly paired LIBS-image fusion strategy for stainless-steel burn-duration recognition. The method retains LIBS as the dominant modality and introduces surface images only as constrained auxiliary evidence. PRISM converts the burn-center image ROI into structured local morphological descriptors through regular sampling and position-aware encoding, whereas UGF adaptively regulates the contribution of image evidence according to modal reliability. In ten repeated experiments, PRISM-UGF achieved an accuracy of 82.67% ± 2.83% and a macro-F1 of 82.26%, outperforming LIBS only (76.93% ± 2.86%) and Image only (69.93% ± 3.63%). These results indicate that PRISM-UGF improves LIBS-based burn-duration recognition by incorporating class-level image information through bounded gated fusion, rather than relying on verified patch-to-spectrum spatial correspondence.