基于改进YOLOv3的中药饮片智能鉴别模型研究  

Intelligent Identification Model of Traditional Chinese Medicine Pieces Based on Improved YOLOv3 Algorithm

作　　者：高爽 周志强 钟思羽 黄显章[1,2] GAO Shuang;ZHOU Zhiqiang;ZHONG Siyu;HUANG Xianzhang(Nanyang Institute of Technology,Henan Key Laboratory of Zhang Zhongjing Formulae and Herbs for Immunoregulation,Nanyang 473004,China;Nanyang Institute of Technology,Henan Engineering Technology Research Center for the Protection and Utilization of Wanyao Traditional Medicine Resources,Nanyang 473004,China;Nanyang Institute of Technology,the Faculty of Computer and Software,Nanyang 473004,China)

机构地区：[1]南阳理工学院河南省张仲景方药与免疫调节重点实验室,南阳473004 [2]南阳理工学院河南省宛药资源保护与利用工程技术研究中心,南阳473004 [3]南阳理工学院计算机与软件学院,南阳473004

出　　处：《世界科学技术-中医药现代化》2025年第2期364-374,共11页Modernization of Traditional Chinese Medicine and Materia Medica-World Science and Technology

基　　金：国家农业农村部现代农业产业技术体系建设专项(CARS-21):山药焦作综合试验站,负责人:黄显章;南阳市科学技术局南阳市科技攻关计划基金项目(KJGG098):基于深度学习的中药饮片鉴别研究,负责人:高爽。

摘　　要：目的针对中药饮片鉴别研究中的饮片漏检、误检、定位不精准、置信度低等问题,通过对小目标和重叠度高的目标具有良好检测效果的YOLOv3算法进行改进,提升中药饮片智能检测识别的准确率。方法采集常见的148种中药饮片图像,构建中药饮片RGB图像数据集。在原始YOLOv3算法模型基础上,通过K-means聚类算法选取合适的锚点框尺寸;引入CIoU损失函数进行边界框回归,提高边界框的定位精度、置信度等;将传统的非极大值抑制NMS改进为DIoUNMS,降低YOLOv3算法对重叠度高的密集目标的漏检、误检等问题。结果对148种中药饮片进行测试,改进后的算法实现了98.47%的平均检测精度均值,相比原始YOLOv3算法提升了1.83%;对密集、重叠度高等复杂情况下的饮片实现了更好的检测效果,饮片漏检、误检、定位不精准、置信度低等问题在一定程度上得到了相应的缓解。结论改进后的算法有效提升了中药饮片的识别精度和泛化能力,为中药饮片实现自动化智能检测提供新的参考。Objective To improve the accuracy of intelligent detection and evaluation of traditional Chinese medicine(TCM)pieces and solve the problems of leakage,misdetection,inaccurate localization and low confidence in the study of TCM pieces identification,YOLOv3 algorithm which has good detection effect for high overlap and small targets was improved.Methods An RGB image database containing 148 commonly used TCM pieces was established.Based on the YOLOv3 algorithm model,the anchor box size was improved by K-means clustering algorithm.The CIoU loss function was introduced for bounding box regression to improve the localization accuracy and confidence of bounding boxes.The traditional non-maximum suppression was improved to DIoUNMS to reduce the problems of missed detection and false detection of dense targets with high overlap by YOLOv3 algorithm.Results 148 kinds of TCM pieces were tested with the improved algorithm,and the average detection accuracy of 98.47% was achieved,which is 1.83% better than the original YOLOv3 algorithm.It realizes better detection effect for TCM pieces in complex situations such as dense,high overlapping,etc.Problems such as leakage,misdetection,imprecise positioning and low confidence level have been alleviated to a certain extent.Conclusion The improved algorithm effectively improves the recognition accuracy and generalization ability of TCM pieces,providing a new reference for the realization of automated intelligent detection of TCM pieces.

关 键 词：中药饮片 深度学习 YOLOv3 损失函数 非极大值抑制 

分 类 号：R282.71[医药卫生—中药学]