Multiscale Spectral–Spatial Fusion for Controlled-Mixture Food Foreign-Object Localization with Conditional Stability Bounds
Localizing small foreign objects in hyperspectral food images requires distinguishing materials while preserving spatial detail. This study presents a multiscale spectral–spatial fusion network with wavelength-aware branches and adaptive fusion weights that incorporate branch sensitivity bounds. A detection-margin analysis accounts for changes in both branch features and fusion weights. The bound uses centered feature disagreement and a finite-radius interaction term to give a sufficient condition for pixel-label stability under bounded perturbations of calibrated reflectance. A predicted foreground component is preserved if the condition holds for its pixels and those of its external background ring. We evaluate localization and prediction stability on controlled mixtures of public fruit recordings and material-library spectra, with prescribed masks, fixed geometry, and bounded reflectance errors.
Authors
- Baoping Wang (ORCID: https://orcid.org/0000-0002-0999-5679)
- Yuheng Wan
- Mingyang Liu
Institutions
- Shenyang Agricultural University (CN)
- Guangdong University of Science and Technology (CN)
Publication Details
- Journal
- Algorithms
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3390/a19111006
- Primary Topic
- Spectroscopy and Chemometric Analyses
- Type
- article
- Field-Weighted Citation Impact
- 0.00