摘要
为解决市民口罩佩戴目标检测中因小尺寸目标较多导致其识别精度不高的问题,提出一种基于YOLOv3改进的算法M_YOLOv3。重构特征金字塔机制,把原本3*3的类金字塔结构扩建为4*4尺寸,把先验框数量由9增加到16,通过以上方法降低神经网络感受野,增强M_YOLOv3对小尺寸目标的敏感度。将原有的损失函数IoU替换为D IoU,解决边框回归时难以确认梯度下降方向的问题。基于网络公开的4065张口罩数据集的实验结果表明,M_YOLOv3的mAP(平均精度均值)为88.4,较Tiny_YOLOv3和YOLOv3的mAP分别提升了15.9和7.2。
To solve the problem that the recognition accuracy is not high enough due to the large number of small-sized targets in the detection of citizens wearing masks,an improved algorithm based on YOLOv3 was proposed and it was called M_YOLOv3.The feature pyramid mechanism was reconstructed,the original 3*3 pyramid like structure was extended to 4*4 size,the number of prior frames was increased from 9 to 16,the receptive field of neural network was reduced and the sensitivity of M_YOLOv3 to small targets was enhanced.The original loss function IoU was replaced with D IoU,which solved the problem that it is difficult to confirm the descent direction of gradient in border regression.Based on the data set of 4065 masks published on the Internet,the mAP of M_YOLOv3 is 88.4,which is higher than that of Tiny_YOLOv3 and YOLOv3 by 15.9 and 7.2,respectively.
作者
曾成
蒋瑜
张尹人
ZENG Cheng;JIANG Yu;ZHANG Yin-ren(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610000,China;Institute of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430081,China)
出处
《计算机工程与设计》
北大核心
2021年第5期1455-1462,共8页
Computer Engineering and Design
