[4Yin-B] Poster 7

[4Yin-B-14] ADAPT+: Partial Test-Time Fine-Tuning for Robust Hyperspectral Image Reconstruction

〇Zhaolu Chen¹, Yi Xu¹, Xian-Hua Han¹ (1. Rikkyo University)

Keywords:

Hyperspectral Image Reconstruction,Test-time Adaptation,Fine-tuning

Hyperspectral images (HSIs) capture rich spectral information through dense sampling across contiguous wavelength bands, typically represented as a three-dimensional (3D) data cube. In modern Coded Aperture Snapshot Spectral Imaging (CASSI) systems, HSIs are acquired in a compressed form as two-dimensional (2D) measurements, necessitating computational reconstruction. Although deep learning models have shown strong potential for addressing this inherently ill-posed inverse problem, their performance often degrades under distribution shifts in illumination conditions and material properties between the training data and real-world scenes. While full-parameter test-time adaptation can alleviate this discrepancy, it incurs prohibitive computational costs and inference latency. To address this issue, we propose ADAPT+, a test-time adaptation framework that performs strategic partial fine-tuning guided by a temporal stability as an early-stopping criteria. Experiments on the KAIST dataset demonstrated that ADAPT+ achieved an average gain of 2.11 dB in PSNR and 0.0080 in SSIM, while only 14.21% of the total parameters are being fine-tuned.