Abstract: In the field of fine-grained image classification of traditional Chinese medicine, the lack of a comprehensive and balanced dataset has been a major obstacle. To advance research in fine-grained image recognition of Chinese herbs, the Herb-150 fine-grained Chinese herb dataset was constructed. To address the issue of deep neural networks easily losing discriminative detailed features in this task, a fine-grained feature-enhanced CHMRN(Chinese herbal medicine recognition Network) was proposed. This network introduces a top-down feature fusion module that integrates multi-scale semantic information, thereby capturing comprehensive contextual features. Additionally, a bottom-up channel information compensation module is designed to enhance the expressive power of fine-grained features, ensuring the accurate capture of subtle differences between traditional Chinese medicine categories. Experimental results show that the CHMRN achieves an accuracy rate of 93.91% on the Herb-150 dataset. Compared to fine-grained image classification models CMAL-Net and PIM,?it achieves an improvement of over 3%, while also outperforming models such as IELT, SR-GNN, I2-HOFI, and SIM-OFE. This significant improvement highlights the effectiveness of the proposed architecture in fine-grained classification tasks. The CHMRN not only improves the accuracy of traditional Chinese medicine identification, but also provides valuable references for similar fine-grained image classification applications.

Key words: Deep learning, Fine-grained image classification, Traditional Chinese medicine, Feature extraction, Feature fusion

摘要： 在传统中草药细粒度图像分类领域，缺乏一个全面且平衡的数据集。为推进中草药细粒度图像识别研究，构建了Herb-150细粒度中草药数据集。针对该任务中深层神经网络易丢失判别性细节特征的问题，提出了细粒度特征增强的CHMRN网络，通过引入自顶向下的特征融合模块整合多尺度语义信息，从而捕捉全面的上下文特征；同时，设计了自底向上的通道信息补偿模块，以增强细粒度特征的表达能力，确保能够准确捕捉中药类别之间的细微差异。实验结果表明，CHMRN网络在Herb-150数据集上的准确率达到了93.91%，相比细粒度图像分类模型CMAL-Net、PIM提升了超过3%，同时优于IELT、SR-GNN、I2-HOFI、SIM-OFE模型，验证了所提架构在细粒度分类任务中的有效性。CHMRN网络不仅提高了传统中药识别的准确性，还能为类似的细粒度图像分类应用提供参考。

关键词: 关键词: 深度学习, 细粒度图像分类, 中草药, 特征提取, 特征融合