基于改进YOLOv5的针灸针小目标检测算法研究  

Acupuncture Needle Small Object Detection Algorithm Based on Improved YOLOv5

作　　者：陆靖桥 万方前 李恒聪 王奕乔 王传池 胡镜清 LU Jingqiao;WAN Fangqian;LI Hengcong;WANG Yiqiao;WANG Chuanchi;HU Jingqing(Xin-Huangpu Joint Innovation Institute of Chinese Medicine,Guangzhou 510000,China;Institute of Basic Theory for Chinese Medicine,China Academy of Chinese Medical Sciences,Beijing 100027,China;Tianjin University of Traditional Chinese Medicine,Tianjin 301617,China;Modern Traditional Chinese Medicine Haihe Laboratory,Tianjin 301617,China)

机构地区：[1]广东省新黄埔中医药联合创新研究院健康产业创新中心,广州510000 [2]中国中医科学院中医基础理论研究所,北京100027 [3]天津中医药大学,天津301617 [4]现代中医药海河实验室,天津301617

出　　处：《世界科学技术-中医药现代化》2025年第1期202-210,共9页Modernization of Traditional Chinese Medicine and Materia Medica-World Science and Technology

基　　金：广州市科技计划项目(2024B03J1239):智慧针灸诊室标准研究及在面瘫治疗中的应用,负责人:高俊虹。

摘　　要：随着针灸学科的科学化、现代化,各种针灸医疗设备不断创新,尤其是智能针灸诊疗设备的出现,针灸临床“留针”阶段的针灸针自动化识别成为需求热点。针对针灸针识别过程的输入图片尺寸过大、针灸针细长、针灸针分布密集等问题,本文提出一个适用于针灸针小目标的改进YOLOv5“留针检测模型”ANODM(Acupuncture Needle Object Detection Model):(1)输入预处理阶段,将图片切分为多个Patch分别预测,以放大小目标尺寸;(2)模型结构层面,在原3个检测层新增一个小目标检测层以提升对小目标的识别能力,将骨干网络的普通卷积替换为空洞卷积以增大感受野,对不同Stage的特征融合;(3)后处理阶段,使用Soft-NMS降低正样本的漏检率,并结合余弦相似度匹配降低负样本的错检率。实验结果表明,对比原YOLOv5算法,本文改进方法在针灸针小目标数据集上的检测精度提升4.2%,能较好地完成针灸针小目标检测任务。With the scientific and modernization of acupuncture,various kinds of acupuncture medical equipment continue to innovate,especially with the emergence of intelligent acupuncture diagnosis and treatment units,automatic detection of acupuncture needles in the"needle retention"stage of acupuncture clinical practice has become a hot demand.Aiming at the problems that the input image size is too large,the acupuncture needles are slender,and the acupuncture needles are densely distributed,the Acupuncture Needle Object Detection Model(ANODM),an improved YOLOv5 model for acupuncture needles,is proposed in this paper.(1)In the preprocessing stage,the image is divided into multiple patches for prediction,respectively.(2)At the model structure level,a new small object detection layer is added to the original three detection layers to improve the recognition ability of small objects.The ordinary convolution of the backbone network is replaced by the dialated convolution to increase the sensitivity field.Features of different stages are fused.(3)In the post-processing stage,Soft-NMS is used to reduce the miss rate of positive samples,and cosine similarity match is used to reduce the error rate of negative samples.The experimental results show that,compared with the original YOLOv5,the detection accuracy of the improved YOLOv5 in this paper is improved by 4.2%on the acupuncture needle small object dataset,which can better complete the acupuncture needle small target detection task.

关 键 词：针灸 针灸针小目标检测 深度学习 YOLOv5 小目标 

分 类 号：R241[医药卫生—中医诊断学]