The quality and flavor of tobacco are of paramount importance, and mold can significantly degrade these attributes. To address this issue, we utilize hyperspectral technology to capture the spectral information of cigarette tobacco across three distinct mold stages: nonmoldy, near-moldy, and moldy. We introduce a novel global-view (GV) attention mechanism, which leverages both 3 × 3 local spatial convolution and global attention to integrate global features effectively. Subsequently, the GV-Net model is designed and implemented for the classification of tobacco mold levels. The model achieves an accuracy of 94.39%, precision of 95.02%, recall of 95.42%, and F1 score of 0.9516. These metrics surpass those of multiple existing attention mechanisms, demonstrating our model’s superior generalization capabilities. The findings of this study not only facilitate the rapid detection and management of mold in tobacco by manufacturers but also help maintain product quality and consumer trust, underscoring our model’s significant practical value and application potential.