摘要
针对单次多重检查(single-shot detector, SSD)方法对小目标物体检测的鲁棒性较差的问题,提出了一种改进的单次多重检查方法应用于电缆隧道明火识别中,通过选取不同层次的特征映射图进行多尺度检测,并引入一组不同大小和长宽比的默认边界框,以增强对小范围明火识别的鲁棒性。首先,构建SSD网络模型,选取6组不同尺寸的特征映射图,并对采集到的明火图像进行预处理;然后,对整个SSD网络进行训练并对默认边界框的大小等参数选取进行优化;最后,得到应用于电缆隧道明火识别的SSD网络,并通过测试集对该方法的识别速度和准确率进行了检验。应用结果表明该方法具有计算速度快、识别准确度高等特点,在电缆隧道明火识别中具有工程应用的参考意义。
Aiming at the problem that the single-shot detector (SSD) method is less robust to small target detection,an improved single-multiple inspection method is proposed for cable tunnel fire detection,multi-scale detection is performed by selecting diferent levels of feature maps and a set of default bounding boxes with different sizes and aspect ratios are introduced to enhance the robustness of identification of small flames.First,to construct the SSD network model,perform multi-scale detection by selecting six different size of the feature maps,and pre-process the collected open flame image.Then train the entire SSD network and optimize parameters such as the size of the default bounding box.Finally,the SSD network applied to the open fire identification of cable tunnel is obtained,and the recognition speed and accuracy of the method are tested by test set.It shows that the method has the characteristics of fast calculation speed and high recognition accuracy.It has the reference meaning of engineering application.
作者
吴宏晓
黄顺涛
崔江静
廖雁群
曾啸
孟安波
WU Hongxiao;HUANG Shuntao;CUI Jianging;LIAO Yanqun;ZENG Xiao;MENG Anbo(Zhuhai Power Supply Bureau of Guangdong Power Grid,Zhuhai Guangdong 519000, China;School of Automation, Guangdong University of Technology, Guangzhou Guangdong 510006,China)
出处
《宁夏电力》
2018年第5期1-5,共5页
Ningxia Electric Power
基金
国家自然科学基金项目(61876040)
广东电网有限责任公司科技项目(GDKJXM20162047)
关键词
电缆隧道
明火识别
深度学习
SSD
cable tunnel
open flame identification
deep learning
SSD
