基于层级识别模型的输电线路杆塔小金具缺陷识别方法

A Defects Recognition Method for Small Fittings in Power Transmission Towers Based on Hierarchical Recognition Model

针对输电线路无人机巡检图像分辨率越来越高,造成的小金具缺陷目标占比小、特征相似度高,进而导致缺陷发现率低和误报高的问题,文章提出一种基于层级识别模型的小金具缺陷识别方法。首先采用级联卷积神经网络(Cascade R-Convolutional Neural Networks,Cascade R-CNN)、更快速卷积神经网络(Faster R-Convolutional Neural Networks,Faster R-CNN)和RetinaNet、YOLOv3算法分别进行算法融合得到高精度的连接金具检测模型和高实时性的小金具识别模型;然后将二者组成层级识别模型,即先对输入图片检测连接金具,再对连接金具识别小金具缺陷。实验结果表明,与直接识别小金具的模型和层级模型中未使用模型融合策略的算法相比,该算法的小金具缺陷识别平均精度均值(mean Average Precision,mAP)和查全率均可达到最高值:0.762和0.826,算法有效性得到有效验证。

Aiming at the increasingly higher resolution of UAV inspection images on transmission lines,the proportion of small fitting defect targets are small,while similar features are relatively high,which leads to low defect discovery rate and high false alarms,a defects recognition method for small fittings based on hierarchical recognition model is proposed in this paper.Firstly,this study created a highprecision detection model for link fittings and a high-real-time recognition model for small fit-tings by respectively fusing Cascade R-CNN and Faster R-CNN,RetinaNet and YOLOv3.Secondly,the two models are combined as a hierarchical recognition model,which firstly detects link fittings in an image and then recognizes defects in small fittings among link fittings.Experimental results show that compared with single models for small fitting recognition and hierarchical models without algorithm fusion,the proposed method can achieve a highest mean Average Precision of 0.762 and a highest recall of 0.826.The effectiveness of algorithm is verified prosperously.