Research And Design Of Aerial Patrol Inspection Image Recognition System For Transmission And Transformation Lines Based On Deep Learning

With the development of new technologies,the use of UAVs to conduct aerial inspections of power transmission and transformation lines has become a new inspection method.From the aerial inspection images,problems in transmission and transformation lines are effectively discovered in time,which is currently a hot spot in research.At present,the main solution is to use deep learning technology through large servers to detect the images collected by the UAV.However,this places high requirements on the detection equipment,and the separation of image acquisition and detection has caused certain delay.Therefore,there is an urgent need for an algorithm that can be loaded in mobile devices carried by drones,which has both excellent detection effect and real-time detection speed.This paper focuses on the deep learning object detection algorithm and conducts related research on aerial inspection images of power transmission and transformation lines.The main research contents of this article are as follows:(1)Through research and analysis of aerial inspection images of power transmission and transformation lines,three types of components including insulators,suspension clamps and anti-vibration hammers were selected for research.Then,from the 10,000 aerial inspection images of power transmission and transformation lines,3822 images that meet the requirements were selected for the labeling of three types of components.Finally,using 3822 aerial inspection images of power transmission and transformation lines and 3822 corresponding annotations,a data set of aerial inspection images of power transmission and transformation lines was constructed,which laid a data foundation for subsequent research.(2)A multi-scale parallel fusion real-time detection algorithm is proposed.First,the algorithm uses a combination of different sizes of inverse residual blocks and different types of activation functions to design a lightweight feature extraction network.Then,in order to improve the detection accuracy,a lightweight parallel fusion structure was designed and applied to multiple scales of the lightweight feature extraction network for prediction.Compared with the existing real-time detection algorithms,the multi-scale parallel fusion real-timedetection algorithm proposed in this paper has better detection accuracy,and has excellent real-time detection performance on mobile devices.(3)Based on the proposed multi-scale parallel fusion real-time detection algorithm,a personalized aerial inspection images recognition system for power transmission and transformation lines was designed and implemented.The user can train a model that meets the requirements by directly interact with the graphical user interface without programming.The aerial inspection image recognition system for transmission and transformation lines can generate the data set needed for model training from the constructed data set and manage it.At the same time,the system supports the selection of training parameters and feeds back the information and training progress in the training process in real time provide dedicated page to verify,detect,and view the test results of the trained model.The system designed and implemented in this paper has tremendous practicality and scalability.