Enhancing Fritillariae Cirrhosae Bulbus authentication via Mask-AIR: a masked autoencoder driven information...

Fritillariae Cirrhosae Bulbus (FCB), a valuable medicinal and edible plant with antitussive properties, is frequently adulterated due to resource scarcity and high market price, necessitating rapid and accurate authentication. However, existing methods (morphological, molecular, chromatographic, etc.) are subjective, destructive, or suffer from DNA degradation. To address these limitations, this paper proposes a multi-modal information reconstruction framework based on a masked autoencoder (MAE) and an iterative attention mechanism that integrates a multi-view fusion strategy using three views (hyperspectral imaging, grayscale images, and pseudo-colour images) to enhance the authenticity identification of FCB. Specifically, multi-view data are fused to enrich information dimensions and thereby improve species recognition accuracy. An MAE operating in the latent feature space reconstructs randomly masked tokens from each modality, which enhances feature representation and mitigates cross-modal interference. Subsequently, hierarchical iterative attention module alternates query roles among modalities to balance fusion. Feature extraction, reconstruction, and classification are trained end-to-end, while inference retains only the encoder and classifier, incurring no extra cost. The experiments on a self-built FCB multi-view dataset show that the proposed method significantly improves classification accuracy across all categories and outperforms both conventional and deep learning baselines. Therefore, this work provides an effective non-destructive tool for FCB authenticity identification and can be extended to the authentication of other valuable functional foods.