基于BIM-Res2Net-ICA的中药饮片图像识别模型  

To Construct an image recognition model for Chinese herbal medicine decoction pieces based on BIM-Res2Net-ICA

作　　者：谷瑞 宋翠玲[1] 李元昊 GU Rui;ONG Cui-ling;LI Yuan-hao(Nanjing University,Nanjing 210003,China;Suzhou Industrial Park Institute Of Services Outsourcing,Suzhou 215123,China;Shanghai University of Traditional Chinese Medicine,Shanghai 200003,China)

机构地区：[1]南京大学,江苏南京210003 [2]苏州工业园区服务外包职业学院,江苏苏州215123 [3]上海中医药大学,上海200003

出　　处：《时珍国医国药》2025年第6期1192-1200,共9页Lishizhen Medicine and Materia Medica Research

基　　金：江苏省高职院校教师专业带头人高端研修项目(2023TDFX010)。

摘　　要：中药饮片是中医药文化的重要载体,针对现有中药饮片识别模型准确率和泛化能力不足的问题,文章从增强多尺度特征适应性和突出重要特征区域两方面入手,提出一种新的中药饮片图像识别模型BIM-Res2Net-ICA。首先,在Res2Net内部构建双向融合策略,以提升不同尺度特征融合的有效性,获取更丰富的特征信息;其次,嵌入改进的坐标注意力机制,在空间和通道两个方向上增强模型对目标的方向感知和坐标信息的捕获能力;最后,结合Focal Loss损失函数有效克服样本分布不均的问题,进一步提高分类的准确性。实验表明,该模型在构建的20类中药饮片数据集上的准确率、精确率、召回率、F1-Score分别为97.48%、96.32%、96.92%和97.59%,与其他先进算法相比,具有更低的计算复杂度和更高的准确率,且模型具有较强的泛化能力,能为各种中药识别场景提供准确的算法支持。Chinese herbal medicine(CHM)decoction pieces are a vital carrier of traditional Chinese medicine(TCM)culture.To address the limitations of existing recognition models in terms of accuracy and generalization ability,this paper proposes a novel image recognition model for CHM decoction pieces named BIM-Res2Net-ICA,by enhancing multi-scale feature adaptability and emphasizing critical feature regions.First,a bidirectional fusion strategy is constructed within Res2Net to improve the effectiveness of multi-scale feature fusion and obtain richer feature information.Second,an improved coordinate attention mechanism is embedded to enhance the model's directional perception of targets and its ability to capture coordinate information in both spatial and channel dimensions.Finally,the Focal Loss function is combined to effectively overcome the problem of uneven sample distribution and further improve the accuracy of classification.Experiments demonstrate that the proposed model achieves an accuracy,precision,recall,and F1-Score of 97.48%,96.32%,96.92%,and 97.59%,respectively on the 20 types of CHM decoction pieces dataset.Compared with other advanced algorithms,the model exhibits lower computational complexity and higher accuracy,along with robust generalization capability,providing reliable algorithm support for various CHM recognition scenarios.

关 键 词：中药饮片 BIM-Res2Net-ICA 多尺度特征 改进的坐标注意力 图像识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]