Transmission Line Target Detection Algorithm Based On Deep Learning

With the development of modern society and the progress of technology,the demand for the safety reliability technology of electric power supply is also increasingly large in our country,so the safety technology of electricity in the leading position in the industry,is more key than any kind in the past.As an important part of the power system,regular inspection of transmission lines is essential.In order to ensure the normal and stable operation of transmission lines,it is necessary to focus on inspecting and maintaining power components such as insulators and shockproof hammers during the inspection of transmission lines.Drone-based inspection of transmission lines will generate a large number of images,and checking anomalies in these images usually requires a large amount of manpower.Therefore,people are increasingly interested in using deep learning-based object detection algorithms to automatically detect these images.However,most of the current target detection algorithms based on computer vision run on computers with high computing power and are difficult to deploy on mobile terminals.In the case of limited computing resources,how to implement online detection on UAV is a challenging problem.In order to solve the above problems,this paper studies the object detection of transmission line inspection images based on deep learning,and the main research content of this paper is as follows:1.Aiming at common power inspection targets such as anti-vibration hammers,bird’s nests,insulators and insulator defective components,this paper collects a large number of transmission line inspection images on the Internet,and enhances the data through brightness,contrast,motion blur,rain,snow and fog for model training and performance testing.2.UAV will encounter a variety of weather conditions during aerial photography,this paper proposes an improved object detection network based on C-YOLOv4,which enhances the sensitivity of the model to detecting target position by introducing a lightweight coordinate attention module in the feature fusion stage;Aiming at the problem of class imbalance in object detection,this paper uses Focal Loss and EIOU to redesign the loss function of the detector,balance the positive and negative samples,and improve the classification and detection accuracy of the model.In this paper,the data set is simulated under severe weather conditions such as rain,snow and fog,and the data is enhanced and fed into the model for training and testing,and the experimental results show that the average accuracy m AP of the model reaches 95.68%,which has good detection performance in the current transmission line image target detection task.3.The target size of different scales in transmission lines usually varies greatly,and the image scale of UAV aerial photography varies greatly.To solve this problem,this paper proposes a lightweight network model Mobile Net-YOLO based on dynamic selection mechanism,which first uses the lightweight network Mobile Netv2 as the backbone of the model,combined with SKNet dynamic selection mechanism,and selects convolution kernels of different sizes for targets of different sizes.At the same time,the ASPP structure is used to enhance the receptive fields of the model to adapt to targets at different scales in aerial images.The ULSAM self-attention mechanism is introduced before the detection head,and different attention maps are learned for each feature map subspace.Finally,the model is trained and tested on the transmission line inspection image dataset in this paper,and the experimental results show that the average detection accuracy m AP of the proposed algorithm reaches 87.52%,the average detection speed reaches 58.21 frames per second,and the model volume size is79.98 MB,which meets the requirements of deployment at the edge mobile end of transmission lines.In order to meet the requirements of lightweight and high precision of power inspection,two improved algorithms,C-YOLOv4 and Mobile Net-YOLO,are proposed on the basis of YOLOv4 algorithm.C-YOLOv4 improves the detection accuracy of the model by embedding the coordinated attention mechanism and optimizing the loss function,and the Mobile Net-YOLO algorithm introduces the Mobilenetv2 and SKNet structures to balance the inference speed,model size and accuracy while optimizing performance,making it easier to deploy the model on embedded devices such as mobile phones.Finally,by comparing the detection results before and after network improvement,and comparing the accuracy with the target detection algorithms such as YOLOv3,YOLOv4,SSD,etc.,the model proposed in this paper takes into account the comprehensive performance advantages of light weight,high recognition accuracy and fast detection speed,and has the characteristics of accurate recognition,good applicability and light weight.Finally,a large number of comparative experiments are performed on the public dataset PASCAL VOC to verify the effectiveness of the proposed algorithm.