摘要
针对施工作业面复杂背景下安全帽检测存在的遮挡,光照多变,且目标尺寸不一等问题,提出一种基于YOLO-v3深度学习网络的改进方法。将原网络模型的52×52检测特征图2倍上采样后,与网络中第二个残差块输出的特征图进行信息融合,再经过一组卷积操作得到一个104×104的新检测特征图。该特征图连同原模型的三个特征级联,构成一个四尺度的目标检测网络模型,改进的检测网络实现了低层特征图的边缘轮廓信息与高层特征图的语义信息的融合。通过实际现场不同工况的测试,实验结果表明,与原YOLO-v3模型相比,改进模型检测的召回率、精确率和F1值分别提高了3.7、0.7以及2.29个百分点,满足施工作业面复杂背景下安全帽检测的准确率要求。
Aiming at the safety helmet detection problems in complex background of construction working surface,such as occlusion,changeable illumination and different target sizes,an improved method based on the deep learning network of YOLO(You Only Look Once)-v3 was proposed.After the 52×52 detection feature map of the original network model was upsampled twice,it was fused with the feature map output by the second residual block in the network,and then a new detection feature map of 104×104 was obtained through a group of convolution operations.It was cascaded with the three feature maps of the original model to form a four-scale target detection network model.The improved detection network realized the fusion of the edge profile information of low-level feature map and the semantic information of high-level feature map.Through the testing of various working conditions in the actual site,the results show that compared with the original model,the improved model improves the recall,precision and F1 value of by 3.70,0.70 and 2.29 percentage points respectively,so that the improved model can meet the accuracyy requirements of safety helmet detection in complex background of construction workingsurface.
作者
杨静
张育飞
毛晓琦
YANG Jing;ZHANG Yufei;MAO Xiaoqi(School of Mechanical and Precision Instrumental Engineering,Xi’an University of Technology,Xi’an Shaanxi 710048,China)
出处
《计算机应用》
CSCD
北大核心
2020年第S02期178-182,共5页
journal of Computer Applications
基金
西安市科技计划项目(2017080CG/RC043(XALG036))。
关键词
施工作业面
YOLO网络
安全帽检测
信息融合
多尺度检测
construction working surface
You Only Look Once(YOLO)network
safety helmet detection
information fusion
multi-scale detection