Removing fringes from hyperspectral images using truncated singular value decomposition

Hyperspectral imaging (HSI) often suffers from optical interference fringes that distort resonant spectral features and obscure spatial information. While various hardware and computational mitigation strategies exist, they often become ineffective when interference frequencies overlap with molecular absorption bands or when fringes are irregularly shaped. In this study, we introduce a straightforward, non-iterative approach to eliminate fringes from infrared (IR) HSI by leveraging the wavelength-dependency of fringe patterns in contrast to fixed resonant features in wavelength-variable images. We use singular value decomposition (SVD) to separate them: resonant features are represented by a few principal singular values (SVs), while optical interferences span a wide range of higher-rank SVs. Then, truncated SVD (TSVD) selects low-rank SVs (resonant) and discards high-rank SVs (fringe). We test the TSVD defringing method on synthetic models of increasing complexity, including uniform films, multi-component patterns, and spatially varying molecular orientation samples. Furthermore, the method's defringing performance is demonstrated on an experimental IR image of a sheared polymer film. Our results confirm that the TSVD method successfully suppresses complex spatial fringes and restores accurate spectral profiles without prior knowledge of the sample, providing a robust, data-driven framework broadly applicable across HSI modalities.