改进的Faster R-CNN方法及其在电缆隧道积水定位识别中的应用  

Improved Faster R-CNN method and its application in recognition of cable tunnel water accumulation

作　　者：崔江静[1] 黄顺涛 仇炜[1] 裴星宇[1] 朱五洲[1] 孟安波[2] CUI Jiangjing;HUANG Shuntao;QIU Wei;PEI Xingyu;ZHU Wuzhou;MENG Anbo(Zhuhai Power Supply Bureau of Guangdong Power Grid,Zhuhai 519000,China;School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东电网公司珠海供电局,广东珠海519000 [2]广东工业大学自动化学院,广东广州510006

出　　处：《电力自动化设备》2019年第7期219-223,共5页Electric Power Automation Equipment

基　　金：广东电网有限责任公司科技项目(GDKJXM20162047)~~

摘　　要：针对电缆隧道内积水的问题,提出了一种改进的基于区域建议的卷积神经网络(FasterR-CNN)方法,并将其应用在电缆隧道积水定位识别中。考虑到Softmax分类方法的正则化参数选取会引起概率计算产生问题,改用支持向量机(SVM)进行图像分类,以增强分类的置信度。使用区域建议网络(RPN)提取隧道积水原图中的区域建议,然后用FastR-CNN检测网络在建议框中进行图像识别、SVM分类和位置精修。实验结果表明,所提方法计算速度快、识别精度高,在实际工程中表现出较高的效率。Aiming at the water accumulation problem in cable tunnel, an improved Faster R-CNN ( Faster Regionbased Convolutional Neural Network) method is proposed and applied in the recognition of cable tunnel water accumulation. Considering that the regularization parameter selection of Softmax may have problems with calculating the probability ,the SVM( Support Vector Machine) is used to classify images to improve the classification accuracy. The RPN ( Region Proposal Network) is used to extract proposals from the original images of cable tunnel water accumulation ,and then the detection network of Fast R-CNN is used to carry out image recognition, SVM classification and location refining. The experiment results show that the proposed algorithm has the advantages of fast calculation speed, high recognition accuracy and high efficiency in practical engineering.

关 键 词：电缆隧道 积水定位 区域建议 卷积神经网络 支持向量机 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]