基于RCNN的无人机巡检图像电力小部件识别研究

发布时间：2018-11-26 12:46

【摘要】：随着无人机(UAV)在电力巡线作业中的应用推广,对无人机巡检图像的信息挖掘或目标识别需求也越来越强烈。传统的电力部件识别流程常使用经典的机器学习算法,如支持向量机(SVM)、随机森林或adaboost,结合梯度、颜色或纹理等浅层特征来对电力部件进行识别,难以充分利用无人机巡检图像的信息,并且难以达到较高的准确率。卷积神经网络(CNN)在目标识别中表现优异,在很多目标识别场景之中成为首选算法。基于区域的卷积神经网络(RCNN)通过使用CNN从图像中提取可能含有目标的区域来检测并识别目标,但是计算复杂,难以满足识别海量电力巡检图片的需求。Fast R-CNN和Faster RCNN利用CNN网络提取图像特征,后接一个区域提议层,优化了提取可能含有目标区域的方式并改进识别目标的分类器,使目标的检测和识别几乎实时。本文详细描述了Faster R-CNN算法流程,并在无人机电力线巡检图像部件检测中使用,然后分别对DPM、SPPnet和Faster R-CNN识别方法进行了对比分析,利用实际采集的电力小部件巡检数据构建的数据集对3种方法进行测试验证,并讨论了不同参数对识别结果的影响。实验结果表明,基于深度学习的识别方法实现电力小部件的识别是可行的,而且利用Faster R-CNN进行多种类别的电力小部件识别定位可以达到每张近80 ms的识别速度和92.7%的准确率。
[Abstract]:With the application of UAV (UAV) in power line inspection, the demand of UAV patrol image information mining or target recognition is becoming more and more intense. Traditional power component recognition processes often use classical machine learning algorithms such as support vector machine (SVM) (SVM), random forest or adaboost, combined with gradient color or texture to identify power components. It is difficult to make full use of the image information of UAV patrol, and it is difficult to achieve high accuracy. Convolutional neural network (CNN) is the best algorithm for target recognition because of its excellent performance in target recognition. The region based convolution neural network (RCNN) detects and recognizes the target by using CNN to extract the region that may contain the target from the image, but the computation is complicated. Fast R-CNN and Faster RCNN use CNN network to extract image features, followed by a regional proposal layer, which optimizes the way of extracting possible target areas and improves the classifier for target recognition. The detection and recognition of target is almost in real time. This paper describes the flow of Faster R-CNN algorithm in detail, and uses it in the detection of UAV power line inspection image components, and then compares and analyzes the DPM,SPPnet and Faster R-CNN recognition methods, respectively. The three methods are tested and verified by the data set constructed from the actual data collected from the patrol inspection of power widget, and the influence of different parameters on the identification results is discussed. The experimental results show that the recognition method based on depth learning is feasible. Moreover, the recognition speed of 80 ms and the accuracy of 92.7% can be achieved by using Faster R-CNN to identify and locate various kinds of power components.
【作者单位】：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;山东鲁能智能技术有限公司;国网山东省电力公司;
【基金】：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”
【分类号】：TM75;TP391.41

【参考文献】