基于YOLOv7⁃tiny的血细胞检测算法  

Blood cell detection algorithm based on YOLOv7⁃tiny

作　　者：叶鑫 钟国韵[1] 刘梅锋[2] YE Xin;ZHONG Guoyun;LIU Meifeng(School of Information Engineering,East China University of Technology,Nanchang 330013,China;School of Mechanical and Electronic Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013 [2]东华理工大学机械与电子工程学院,江西南昌330013

出　　处：《现代电子技术》2025年第7期119-125,共7页Modern Electronics Technique

基　　金：国家自然科学基金项目(62162002);江西省主要学科学术和技术带头人领军人才项目(20225BCJ22004)。

摘　　要：血常规检验作为医疗诊断的一项重要方法,主要是对血液中血小板、白细胞和红细胞进行识别和计数。针对血细胞检测存在细胞形状不规则、目标尺度变化大、细胞相互遮挡等问题,提出一种改进YOLOv7⁃tiny的血细胞检测算法——EMCDModel。首先,采用可变形卷积(DCNv3)替换高效长程聚合网络的二维卷积,提出ELAN⁃DF模块,提高了不规则目标特征学习能力,降低了模型参数量和计算量;其次,采用MPDIoU替换原始的CIoU适应血细胞密集分布下的尺度变化,降低其相互遮挡导致的漏检率;在主干加入CBAM注意力机制加强对血细胞关键信息的学习,提高对血小板等小目标的检测精度;最后,通过轻量级上采样算子CARAFE替换颈部网络的最近邻插值法,强化颈部网络的特征融合能力,同时降低模型参数量。在BCCD数据集上的实验结果表明,EMCDModel的平均精度均值(mAP)达到92.8%,模型大小仅有5.5 MB,相较于YOLOv7⁃tiny算法,mAP提升了3.8%,参数量降低了8.15%,有效提升了血细胞检测精度。The blood routine examination is vital for medical diagnosis,identifying and counting components like red blood cells(RBCs),white blood cells(WBCs)and platelets(PLTs).In view of the irregular cell shape,large change of object scale and mutual occlusion of cells in blood cell detection,a blood cell detection algorithm EMCDModel based on improved YOLOv7⁃tiny is proposed.Firstly,the deformable convolutional networks(DCNv3)is used to replace the 2D convolution of the efficient long⁃distance aggregation network,and an ELAN⁃DF module is proposed,which improves the learning ability of irregular object fea⁃tures and reduces the model parameters and computation.MPDIoU(minimum point distance based IoU)is used to replace the original CIoU(complete intersection over union)to adapt to the scale variations in the case of dense blood cell distribution,lowering missed detection rates due to occlusion.The CBAM(convolutional block attention module)attention mechanism is introduced to the backbone to enhance the learning for the key information of blood cells,so as to improve the detection accuracy of small ob⁃jects like PLTs.Finally,the lightweight upsampling operator CARAFE(content⁃aware reassembly of features)is used to replace the nearest neighbor interpolation,so as to improve the feature fusion in the neck network and reduce the model parameters.Tests on the BCCD(blood cell count and detection)dataset show that the EMCDModel achieves an mAP(mean average preci⁃sion)of 92.8%with a model size of 5.5 MB.In comparison with the YOLOv7⁃tiny,the EMCDModel improves mAP by 3.8%,re⁃duces parameters by 8.15%,and improves blood cell detection accuracy effectively.

关 键 词：深度学习 血细胞检测 YOLOv7⁃tiny 注意力机制 可变形卷积 小目标检测 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]