CNN-Based Classification of Ziziphus Seeds with Focal Loss for Overcoming Size-Based Shortcut Learning

Herbal medicines represent a significant global market, yet food safety remains threatened by counterfeit products morphologically resembling authentic samples. Models trained on limited datasets are prone to shortcut learning, relying on superficial features rather than intrinsic morphological characteristics. This study identified size-based shortcut learning as a critical factor degrading the classification of Ziziphus jujuba Mill. var. spinosa and its counterfeit Ziziphus mauritiana Lam., and demonstrated that focal loss alone can effectively mitigate this issue. Models trained on the internal dataset were evaluated on an external dataset acquired with the Herb-X. On the internal test set, all configurations achieved high classification accuracies (≥98%), thereby obscuring meaningful differences in external generalization. However, consistent performance degradation was observed on the external dataset. The cross-entropy model trained on background-removed data dropped to 82.08 ± 10.97%, while size-normalized models recovered to 84.17 ± 10.15% (upsizing) and 88.94 ± 6.76% (downsizing), confirming that suppressing size shortcuts improves external generalization. The focal loss model, without any preprocessing, achieved 90.88 ± 2.71%, reducing the internal-external generalization gap from 16.18 to 8.11 percentage points. Grad-CAM++ and loss analyses confirmed that the focal loss model attended to intrinsic morphological features rather than object size. This study provides a practical, preprocessing-free approach for reliable herbal-medicine authentication in field conditions.