基于改进Yolov3半月板撕裂的自动检测及分类  

Meniscal tear automatic detection and classification based on improved Yolov3

作　　者：王瑞宇 王娟[1,2] WANG Ruiyu;WANG Juan(Tianjin Key Laboratory of Complex System Control Theory and Application,Tianjin University of Technology,Tianjin 300384,China;School of Electrical Engineering and Automation,Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津理工大学天津市复杂系统控制理论及应用重点实验室,天津300384 [2]天津理工大学电气工程与自动化学院,天津300384

出　　处：《天津理工大学学报》2024年第5期114-118,共5页Journal of Tianjin University of Technology

基　　金：国家自然科学基金(61773286)。

摘　　要：针对医师进行半月板撕裂诊断时,存在半月板区域面积小,无法有效识别的问题,提出一种改进后Yolov3的定位裁剪方法。用高效通道注意力网络(efficient channel attention network,ECANet)作为移动翻转瓶颈卷积(mobile inverted bottleneck convolution,MBConv)中的注意力机制模块,命名为MBConv+。MBConv+模块堆叠构成EfficientNet-b0+,作为Yolov3的特征提取网络,对半月板进行检测及定位裁剪,并采用支持向量机(support vector machines,SVM)进行分类实验。实验结果表明:改进后Yolov3算法与原算法相比,在平均精确率均值(mean average precision,mAP)相近条件下,训练时间大幅缩减,帧数(frames per second,FPS)提升了1.3221,达到了29.0691。与原始数据集相比,裁剪后图像在分类性能上有提升,精度提升了7.739%,达到91.251%。An improved positioning and cropping method,Yolov3,is proposed to solve the problem that the meniscal area is small and can not be effectively recognized during meniscal tear diagnosis.ECANet is used as the attention mechanism module in MBConv,named MBConv+,which stacks to constitute EfficientNet-b0+as the feature extract network of Yolov3 for the detecting,positioning and cropping of the meniscus,and the classification experiments are performed with a support vector machine(SVM).The experimental results show that the improved Yolov3 algorithm significantly reduced the training time compared with the original algorithm under similar mAP conditions,and the frames per second(FPS)increased by 1.3221 and reached to 29.0691;compared with the original dataset,the cropped images improved in classification performance and the accuracy increased by 7.739%to 91.251%.

关 键 词：半月板撕裂 目标检测 Yolov3算法 图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]