Boosted Spectral Perception via Synergistic Vision for Cross-Domain Few-Shot Hyperspectral Image Classification

Cross-domain few-shot learning (CDFSL) hyper-spectral image classification (HSIC) techniques have significantly enhanced the performance of HSIC under a few annotations. Due to the scattered spatial distribution of HSI and limited labels, the spatial information of the HSI sample is restricted, which means the spectral information may contain more discriminative category features than spatial information. However, most existing CDFSL HSIC methods lack targeted spectral discriminative feature extraction and struggle to capture spectral subtleties. To address this, a Human Eye-like Dual-Branch Network (HEDB-Net) is proposed, comprising a central branch and a peripheral branch. The central branch focuses on extracting core features from complete spectral dimension HSI patches. Simultaneously, these patches are input into the peripheral branch via random sampling and reintegration of spectral bands, enabling the model to perceive initially dispersed spectral information. These two branches represent the model’s central and peripheral vision for the spectrum, allowing it to focus on different levels of feature information and significantly boosting spectral perception capabilities. To fully leverage the advantages of synergistic vision, a consistency constraint is applied to align their category probability prediction distributions, which helps ensure consistency in classification decisions between the two branches. Experimental results on the PaviaU dataset demonstrate that HEDBNet outperforms other advanced algorithms in terms of overall accuracy(OA), average accuracy(AA), and Kappa coefficient. This proves that HEDBNet can simultaneously focus on primary and secondary spectral features, thus significantly boosting classification accuracy.