改进YOLOv7算法的龋齿图像检测  

Improved YOLOv7 algorithm for caries image detection

作　　者：范晓聪 姚竟发 滕桂法[1] 马永平 FAN Xiaocong;YAO Jingfa;TENG Guifa;MA Yongping(School of Information Science and Technology,Hebei Agricultural University,Baoding 071000,China;Department of Software Engineering,Hebei Software Institute,Baoding 071000,China;Stomatology Department of Baoding No.2 Hospital of Hebei Province,Baoding 071000,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071000 [2]河北软件职业技术学院软件工程系,河北保定071000 [3]河北省保定市第二医院口腔科,河北保定071000

出　　处：《现代电子技术》2024年第17期79-87,共9页Modern Electronics Technique

基　　金：国家自然科学基金项目(U20A20180);中国高校产学研创新基金资助课题(2021LDA10005);河北省重点研发计划项目(21327405D)。

摘　　要：针对口腔医疗资源紧缺和龋齿治疗效率不足的问题,提出一种改进YOLOv7的龋齿图像检测算法,旨在协助医生进行更有效的医疗诊断,同时增强患者对预防龋齿的意识。首先,在YOLOv7算法的主干网络引入ECA-MobileOne网络模块代替原有的ELAN模块,降低模型参数量,提高对小目标龋齿特征的有效提取;其次,在特征图输出层采用自适应特征融合(ASFF),自适应地学习各尺度特征图在融合时的空间权重,充分获取口腔图像中不同尺度下的关键特征,提高检测的全局性和准确性;另外,采用soft-NMS算法替换原有的非极大值抑制算法(NMS),在牙齿异位或重叠等情况下能更有效地提升检测效果。使用在保定市第二医院口腔科采集的口腔照片数据集进行实验,结果显示,改进后的算法mAP达到93.4%,相较于原始YOLOv7算法提高了5.5%,并且与当前主流算法相比,具有一定的先进性,为促进口腔健康的整体改善提供了新的技术支持。In view of the shortage of oral healthcare resources and the inefficiency in caries treatment,a scheme of an improved YOLOv7 algorithm for caries image detection is proposed to assist doctors in making more effective medical diagnosis,and assist patients in enhancing their awareness of caries prevention.The network module ECA-MobileOne is introduced into the backbone layer of YOLOv7 algorithm instead of the original module ELAN,so as to reduce the number of model parameters and improve the feature extraction of little caries(little objects).An adaptive spatial feature fusion(ASFF)is used in the output layer of feature map to adaptively learn the spatial weight of the feature maps of different scales in the process of fusion,and fully acquire the key features of different scales of the oral cavity image,so as to improve the global nature and accuracy of the detection.In addition,the soft-NMS algorithm is used to replace the original NMS(non-maximum suppression)algorithm,so as to improve the detection effect more effectively in the case of teeth ectopic or overlapping.Experiments were conducted based on an oral photograph dataset collected from the stomatology department of Baoding No.2 Hospital.The results show that the improved algorithm achieved an mAP(mean average precision)of 93.4%,which was 5.5%higher in comparison with that of the original YOLOv7 algorithm.It can be seen that the improved algorithm is advanced in comparison with the current mainstream algorithms,and can provide technical support for the overall improvement of oral health.

关 键 词：龋齿检测 MobileOne 自适应特征融合 YOLOv7 soft-NMS 图像检测 

分 类 号：TN911.1-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]