Bird's Nest Identification In Aerial Image Of High-voltage Power Line Inspection

In recent years,high-voltage overhead transmission lines have been widely distributed in China,and many lines are located in complex mountainous areas.In this environment,birds often nest on the towers and insulators of transmission lines,seriously impacting the safe operation of the grid.In the past,manual patrol inspection was mainly used in China's railway patrol inspection.But with the development of technology,drone inspection technology emerges as the times require.Unmanned aerial vehicle patrol process will take a lot of images.However,the detection of the bird's nest in the State Grid mainly relies on drone inspections,whose judgments and labels of images need manual analysis,and the manual detection method is time-consuming and laborious.Therefore,it is necessary to study the automatic detection technology of bird nest identification.Aiming at the problem of bird's nest recognition in aerial images of high voltage power lines,this paper adopts deep learning technology to achieve high accuracy and high efficiency automatic detection of bird's nest recognition in aerial images.The image data set of high-tension transmission line with bird's nest is constructed in this paper,and we label and classify image data sets.By comparing various image recognition algorithms,we choose YOLOv3 algorithm and improve it.The main contributions of this paper are as follows:1.The first part builds image data sets and clusters the label frame of the image in the data set.Before the improvement of YOLOv3,anchor in YOLOv3 was clustered according to COCO data set.But the big difference between the anchor and n est data set will slow down the fitting speed of network.K-means clustering of the tag box of nest data sets can get anchor which is closer to the size of the object in the sample,which can reduce the difficulty of fine-tuning the anchor to the actual location of the network and improve the efficiency of network fitting.2.The loss function of YOLOv3 algorithm is improved.The loss function of the unmodified YOLOv3 algorithm will lead to the cost of the overall loss caused by the error of the large target is much less than that of the small target,so that the accuracy of small target recognition is not high.At the appearance of targets in var ious size in the data set,the width of the label box in the original loss function is normalized,such that the network treat the large and small targets equally.Compared with the loss function of the unimproved YOLOv3 algorithm,the loss function can be suitable for all kinds of large and small targets after being improved.3.Improvement of class imbalance in data sets.A large number of backgrounds where there are only 1.6 nests per image in the nest data sets on average,which leads to the imbalance of target and background categories.In this paper,Focal loss algorithm is added to YOLOv3 loss function.By adding Focal loss algorithm,it is found that not only can the network focus more on the training of bird's nest,and improve the accuracy of bird's nest recognition,but also can improve the detection speed.4.We improve of connection layers of feature pyramid network.In order to further improve the accuracy of YOLOv3 algorithm for small targets recognition.In this paper,two schemes are proposed to modify the number of connection layers of the characteristic pyramid network in YOLOv3 algorithm.A comparison of the two modification schemes indicates that reasonable revisions of the structure through combining the low-rate and strong-semantic features with high-rate and weak-semantic features,the recognition of small target be improved without cutting down the accuracy of large target.5.Finally,this project gives improvement of convolution layer.Network weights have been trained through the unimproved YOLOv3 algorithm are too large,which is not suitable for real-time operation on UAV.For the sake of running the recognition network on UAV in real time,it is necessary to further reduce the capacity of network weight without reducing the recognition accuracy.In YOLOv3 algorithm,the convolution layer based on ResNet network is changed into the convoluti on layer based on DenseNet network.The modified algorithm training model can improve the recognition accuracy and reduce the weight when the detection time increases slightly.In addition,this paper finds out that it can still achieve high accuracy witho ut pre-training weight.The experimental results show that:(1)The maximum average detection accuracy(map)and detection time obtained by clustering experiments are consistent with YOLOv3 algorithm.The map graph after clustering can reach the maximum faster and more stable than YOLOv3 algorithm;(2)The map obtained by the experiment after the improvement of loss function is increased from 87.8% to 88.6%,the F1 value is increased from 0.89 to 0.91,whose training time is shortened by 2.8%;(3)Since the imbalance of data sets get improvement,the maps obtained from experiments increased to 89.5%,whose F1 value increased to 0.92,and the detection time and training time were reduced by 9.7% and increased by 7.1% respectively;(4)On the basis of the change of the number of pyramid network connection layers,the maps increased to 90.8%,the F1 value increased to 0.96,the detection time increased 21.4%,and the training time increased 2.7%;(5)Owning to the improvement of convolution layer,the maps received from the result increased to 91.1%,the F1 value increased to 0.97,the weight reduced by 40%,the detection time increased by 17.6%,the training time increased by 23.4%.On the condition of pre-training weight the map was 90.8% and the F1 value increased to 0.95.Contrast to the YOLOv3 algorithm before the improvement,other performance indicators from new improved algorithm are better than the original algorithm's in addition to a slight increase of 25% in detection time.It can be seen that the impr oved algorithm can be better adapted to the bird's nest recognition in aerial images,and has important value for the application of UAV in power inspection.