Toxic chinese herbal medicine recognition in real-world images via multi-scale and attention-enhanced EfficientNetV2.

Accurate identification of toxic Chinese herbal medicines is critical for public health and clinical safety. However, real-world herbal images often exhibit complex backgrounds and small, indistinct target regions, posing substantial challenges to automated classification systems. In this study, we present a novel image dataset comprising over 4,000 samples from 47 toxic herb categories, captured under diverse environmental conditions to reflect real-world variability. We benchmark several state-of-the-art convolutional neural networks, including ResNet, ResNeXt, and EfficientNet variants, and identify EfficientNetV2 as the most effective baseline. To further enhance model robustness and discriminative capability, we propose an improved EfficientNetV2 architecture incorporating two lightweight yet effective modules: a Multi-Scale Feature Fusion (MSFF) module to integrate hierarchical features, and a Convolutional Block Attention Module (CBAM) to refine both spatial and channel-wise representations. Experimental results demonstrate that our enhanced model achieves 91.28% Top-1 accuracy, 97.52% Top-5 accuracy, and a 90.27% macro F1-score, significantly outperforming baseline methods. Ablation studies confirm the complementary benefits of MSFF and CBAM, and targeted evaluations on challenging image subsets reveal improved resilience to background clutter and small object localization. The proposed architecture offers a high-accuracy, generalizable, and computationally efficient solution for toxic herbal medicine classification and provides a valuable reference for intelligent traditional medicine recognition applications. The GitHub repository for this project is available at: https://github.com/zhuguohua1992/Toxic-chinese-herbal-medicine-recognition-via-enhanced-EfficientNetV2.