目的 解决中药饮片种类繁多、形态相似，人工识别耗时费力且易出错的问题。方法 构建了包含 201 类中药饮片的数据集，并提出了一种轻量化改进的 YOLOv8 算法，具体改进包括在 YOLOv8n 网络中引入 GhostC2f 模块以降低模型参数量， 采用 DySnakeC2f 模块以增强对纤细结构的灵敏度，替换主干网络的池化层为 SimSPPF 模块以加快推理速度，并加入坐标注意力（ coordinate attention， CA）机制以增强对小尺寸目标的特征提取。结果 改进后的算法跨阈值平均精度（ 50%～95%）达到 84.16%，较之前提高了 4.39%，同时模型参数量减少了约 15%。改进的模型成功部署在电脑客户端和手机 APP中，构建了中药饮片自动化识别标注系统。结论 改进后的模型能够有效识别中药饮片，同时系统支持自动数据扩充和升级，从而为中药饮片的快速、准确识别提供了一种新方法。

Objective Addressing the issues of the diverse types and similar shapes of traditional Chinese medicine decoction pieces (TCMDPs), which could make manual identification time-consuming, labor-intensive, and prone to errors. Methods A dataset containing 201 classes of TCMDPs was constructed, and a lightweight improved YOLOv8 algorithm was proposed. The specific improvements include introducing the GhostC2f module in the YOLOv8n network to reduce model parameters, adopting the DySnakeC2f module to enhance sensitivity to fine structures, replacing the pooling layers of the backbone network with SimSPPF to accelerate inference speed, and incorporating the coordinate attention (CA) mechanism to improve feature extraction for small-sized targets. Results The improved algorithm achieved a cross-threshold mean average precision (50%—95%) of 84.16%, representing an increase of 4.39% compared to the previous version, while reducing the model’s parameter count by approximately 15%. The enhanced model was successfully deployed on both computer clients and mobile apps, creating an automated recognition and annotation system for TCMDPs. Conclusion The improved model effectively identifies TCMDPs, while the system supports automatic data expansion and upgrades, providing a novel approach for rapid and accurate identification of TCMDPs.

TP18;R282.710.3

中国中医科学院科技创新工程项目(CI2023E002);国家自然科学基金青年科学基金项目(32202415);山东省研究生优质教育教学资源项目(SDYAL2022041)