Research On High-voltage Tower Bird’s Nest Detection Algorithm Based On Convolutional Neural Network

The detection and removal of bird’s nests in high-voltage towers of transmission lines is extremely important to ensure the stable operation of the power system.At present,there are two problems with the automatic detection of bird nests in transmission line UAV inspection.One is that the background of bird nest images of high-voltage towers is complex,the target area to be detected is small and easily partially obscured by steel towers,and there is a certain amount of leakage in the detection process.The other is that due to the limited performance of mobile and embedded devices,the current target detection model with better accuracy cannot be transplanted to UAV airborne edge devices due to excessive computation,and cannot meet the demand for rapid bird nest detection.To address the above problems,this paper proposes a convolutional neural network-based bird nest detection algorithm for high-voltage poles,improves the commonly used target detection framework YOLOv5 to improve the detection accuracy of bird nests,and lightens the improved YOLOv5 to reduce the computational effort and save bird nest detection time.The research of this paper includes:(1)A method to improve YOLOv5 s is proposed for the problem of small target area and difficult detection in the detection of bird nests in high-voltage towers.The method uses the Kmeans++ algorithm to design anchor frames to obtain nine anchor frames suitable for highvoltage tower bird nest identification and assign them in three detection layers;uses the DeepRFB perceptual field module to replace the SPPF module(spatial pyramid pooling)to obtain a larger perceptual field and reduce the loss of small target information by introducing the convolution of voids with different expansion rates;uses the Focal-EIOU loss function to replace the CIOU loss function of YOLOv5 s,which improves the model convergence ability.(2)A simple YOLOv5 s model based on neck network optimization is designed for problems such as few information features in images and complex image backgrounds.The model optimizes the neck network of the improved YOLOv5 s,and uses the GSConv module to replace the Conv module of the neck network,which improves the detection accuracy while reducing the computational effort of the model.The structure of the C3-X(False)module of the neck network is also improved,based on the reconfiguration of the CA attention module into the C3 f CA convolution module,which further enhances the fusion of position information and improves the detection accuracy of the model.(3)A distillation-based YOLOv5 s model based on knowledge distillation is designed to address the problem that the target detection model is too computationally intensive to be deployed on airborne edge devices.A distilled YOLOv5 s model is obtained by using knowledge distillation to lighten the simple YOLOv5 s model.First,YOLOv5 n is used as the student model and the simple YOLOv5 s model is used as the teacher model.Then,the rich information is transferred to the student model through the soft labels in the teacher model.Finally,a border regression loss strategy is added.Experimental results show that the distilled YOLOv5 s model,which has been lightened,effectively reduces the computational effort and model size with slightly reduced accuracy.To further validate the effectiveness of the proposed method in this paper,the distillation-type YOLOv5 s model was deployed on the Jetson Nano mobile terminal,and good experimental results were achieved.Based on the above work,this thesis researches and implements a high precision and lightweight bird’s nest detection algorithm for high voltage towers.These works have practical significance for the new generation of transmission line inspection drones to realize inspection while patrolling.