摘要
针对YOLOv3目标检测算法存在网络参数量大、检测精度不够高的问题,首先,使用一种轻量化YOLOv3特征提取网络的方法,降低特征提取网络部分的参数量;其次,提出一种多级特征融合网络结构,提高YOLOv3算法特征层的检测效果;最后,采用一种软化的非极大值抑制(soft non-maximun suppression,Soft-NMS)算法,在检测阶段有效避免重叠目标下的漏检。结果表明,相比于YOLOv3算法,改进YOLOv3算法的参数量降低了46%,模型大小约为原模型的50%,在PASCAL VOC 2007数据集上的平均精度均值(mean average precision,mAP)提升了3.5%。
To overcome the defects of YOLOv3 algorithm’s large amount of network parameters and low detection accuracy,firstly,a light-weight YOLOv3 feature extraction network method was used to reduce the amount of parameters in the feature extraction network part.Secondly,a multi-level feature fusion network structure was proposed to improve the detection effect of the YOLOv3.Finally,a soft non-maximun suppression(Soft-NMS)algorithm was used in the detection stage to effectively avoid missed detections.Compared with the traditional YOLOv3 object detection algorithm,the experimental results of the improved YOLOv3 algorithm show that the amount of parameters are reduced by 46%and the model size is about 50%of the original model,with the mean average precision(mAP)on the PASCAL VOC 2007 data set increased by 3.5%.
作者
曹春键
臧强
王泽嘉
屠壮
CAO Chunjian;ZANG Qiang;WANG Zejia;TU Zhuang(School of Automation, Nanjing University of Information Science and Technology, Nanjing 210044, China;Jiangsu Collaborative Innovation Center on Atmospheric Environment and Equipment Technology, Nanjing 210044, China)
出处
《中国科技论文》
CAS
北大核心
2021年第11期1195-1201,共7页
China Sciencepaper
基金
国家自然科学基金资助项目(61973170,51575283)
国家重点研发计划项目(2017YFD0701201-02)
南京信息工程大学大学生创新创业训练计划项目(201910300308,202010300110Y)。
关键词
目标检测
YOLOv3算法
特征融合
非极大值抑制
平均精度均值
object detection
YOLOv3 algorithm
feature fusion
non-maximun suppression(NMS)
mean average precision(mAP)
