基于YOLOv7网络的高分辨率结核杆菌识别方法研究  

Research on High-resolution Mycobacterium Tuberculosis Identification Method Based on YOLOv7 Network

作　　者：刘三雄 LIU Sanxiong(College of Physics and Electronic Engineering,Shanxi University,Taiyuan,030006,China)

机构地区：[1]山西大学物理与电子工程学院,太原030006

出　　处：《网络新媒体技术》2024年第3期26-35,共10页Network New Media Technology

基　　金：国家重点研发计划(编号:2022ZD0118300)。

摘　　要：结核病已经被列为全球十大主要死亡原因之一,针对痰涂片图像背景复杂,结核杆菌目标尺寸小等难题,提出基于痰涂片图像的单阶段YOLOv7结核杆菌检测网络。骨干网络中,分别在不同尺度的高效聚合网络后嵌入CBAM注意力模块,以提取不同尺度的特征,在颈部网络中引入分离合并操作以改进MPConv结构,减少深层网络因卷积下采样引起的图像特征损失,并在头部网络中引进高斯分布距离来避免因边界框重叠而引起的小目标漏检,得到一种精度高、检测速度快的结核杆菌方法。实验表明,改进模型的均值平均精度达到了87.8%,相比基线网络模型提升5.1%,且优于同类算法,对推进结核杆菌的智能化检测具有重要意义。Tuberculosis is one of the ten leading causes of death worldwide.To address the complex background and small target size of Mycobacterium tuberculosis in sputum smear images,in this paper,we put forward a YOLOv7 detection network for detecting Mycobacterium tuberculosis using sputum smear images.In the backbone network,CBAM attention modules are embedded after efficient aggregation networks at different scales to extract features at different scales.We introduce separation and merging operations in neck net-works to improve MPConv structure and reduce image feature loss due to convolutional downsampling in deep networks,and introduce Gaussian distribution distances in the head mesh to avoid small target misses due to overlapping bounding boxes.This technique is exceedingly precise and rapid,rendering it ideal for identifying Mycobacterium tuberculosis.Experiments indicate that the mean average accuracy achieved by the improved model is 87.8%,a 5.1%enhancement compared to the baseline network model,and better than similar algorithms,which is important for advancing intelligent detection of Mycobacterium tuberculosis.

关 键 词：结核杆菌检测 YOLOv7 注意力机制 分离合并操作 高斯分布 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] R52[自动化与计算机技术—计算机科学与技术] TP183[医药卫生—内科学] R446.5[医药卫生—临床医学]