• 遥感科学与应用技术 •

### 基于RCNN的无人机巡检图像电力小部件识别研究

1. 1. 国网山东省电力公司电力科学研究院 国家电网公司电力机器人技术实验室,济南 250002
2. 山东鲁能智能技术有限公司,济南 250101
3. 国网山东省电力公司,济南 250000
• 收稿日期:2016-03-17 修回日期:2016-06-21 出版日期:2017-02-28 发布日期:2017-02-17
• 作者简介:

作者简介：王万国（1984-）,男,硕士,中级工程师,主要从事基于数字图像的变电站和输电线路设备识别、目标跟踪、图像去雾等研究。E-mail: wangwanguo03@qq.com

• 基金资助:
2014年国家电网公司发展项目“无人机巡检实用化关键技术及检测体系研究”

### Study on the Electrical Devices Detection in UAV Images based on RegionBased Convolutional Neural Networks

WANG Wanguo1,2,*(), TIAN Bing3, LIU Yue1,2, LIU Liang1,2, LI Jianxiang1,2

1. 1. Electric Power Robotics Laboratory of SGCC, Shandong Electric Power Research Institute, Jinan, 250002, China
2. Shandong Luneng Intelligence Technology Co., Ltd. Jinan, 250101, China
3. State Grid Shandong Electric Power Company, Jinan, 250000, China
• Received:2016-03-17 Revised:2016-06-21 Online:2017-02-28 Published:2017-02-17
• Contact: WANG Wanguo E-mail:wangwanguo03@qq.com

Abstract:

With the wide application of Unmanned Aerial Vehicle （UAV) in the inspection of power transmission line, the demand for objects detection and data mining from images acquired by UAV also grows significantly. Traditional detecting methods use some classical machine learning algorithms, such as support vector machine （SVM), random forest or adaboost etc. and combine the low level features such as gradient, colors or texture to detect electrical devices. These image features must be carefully designed and changed a lot from various object kinds. Thus, they are not suitable for UAV images with complex background and multiple kinds of object. On the other hand, the disadvantages of these methods are that they cannot take advantage of the high quantity and large coverage of UAV acquired images, and cannot get a satisfactory accuracy. The recent developing Deep Learning method brings light to this problem. Convolutional neural network （CNN) performs excellently in object recognition area and outstand many other methods used in the past. Without the need of extracting images’ features, CNN becomes the many state-of-the-art methods in object recognition rapidly. In object detection, Region-based convolutional neural networks （RCNN) retrieves the region that may contain the object from the images to detect and recognize the object. However, the computation is so expensive that it cannot meet the requirement of detecting massive UAV’s images and cannot be used in practical projects. Fast R-CNN and Faster R-CNN solve this problem by changing the way of object retrieval. They use features produced by CNN network layers and apply a region proposal network layer behind to locate the object. After that, fully connected layers and softmax layer follow to classify the features corresponding to object into special kinds. Using this strategy, Fast R-CNN and Faster R-CNN save lots of time to produce region proposal and can perform object detection at nearly real time. The principle and processes of Faster R-CNN and several other object detection methods are described in this paper, and they are tested for electrical devices detection from images of the power transmission line obtained by UAV. We analyzed the influence of several key parameters to the device detection results, such as the dropout ratio, non-maximum suppression （nms) and batch size. Then, we gave some constructive advice of tun ing parameters in Faster R-CNN. We also analyzed the advantages and weakness of three advanced detection algorithms, including Deformable Part Models （DPM) and two deep learning-based methods named Spatial pyramid pooling networks （SPPNet) and Faster R-CNN. Finally, we constructed image datasets of power transmission line inspection obtained by UAV and tested the three methods above. The recall ratio and accuracy ratio of them are compared and the superiority of the Faster R-CNN is validated. Testing results showed that Faster R-CNN method can detect various electrical devices of different categories in one image simultaneously within 80 milliseconds and achieve an accuracy of 92.7% on a standard test set, which is of great significance in real-time power transmission line inspection. These results also showed the advantages of the Faster R-CNN and we apply Faster R-CNN in our practical projects to detect electrical devices.