Manuscript received October 13, 2025; revised November 12, 2025; accepted December 29, 2025; published May 27, 2026.

Abstract—Hyperspectral Imaging (HSI) captures rich spectral details far beyond RGB cameras. To enable real-time acquisition, contemporary HSI systems often rely on compressive sensing theory, making the reconstruction of hyperspectral data from compressed measurements a fundamental and challenging task in computational imaging. While supervised deep learning has achieved remarkable success in Hyperspectral (HS) image reconstruction tasks, the generalization capability of these existing models is often hindered by limited training data, resulting in poor performance on unseen scenes. To address this issue, we proposed ADAPT (ADAptive Prior Transfer), a test-time prior transfer framework that bridges the gap between the general priors learned by a pre-trained model and the specific information of a target scene. Our method leverages a arbitrary pre-trained deep unfolding or end-to-end network and test-time fine-tuning using a loss function derived from the physical Coded Aperture Snapshot Spectral Imaging (CASSI) imaging forward model and the emprical prior of HSI. Experiments on the KAIST dataset demonstrated that our method improved the reconstruction quality for both transformer- and deep unfolding-based models.

Keywords—hyperspectral image reconstruction, test-time adaptation, self-supervised learning

Cite: Zhaolu Chen and Xian-Hua Han "ADAPT: General-to-Specific ADAptive Prior Transfer for Hyperspectral Image Reconstruction," Journal of Image and Graphics, Vol. 14, No. 3, pp. 366-375, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).