Research And Implementation Of Object Detection Algorithm For Grid Transmission Line Components Recognition

For a long time,object detection has been a very important task in the field of computer vision.It combines techniques such as pattern recognition,deep learning,and image processing.It is also the basis for other higher-level vision tasks such as autonomous driving and instance segmentation.With the rise of deep learning and neural network models,and the improvement of large-scale computer vision data sets,object detection based on deep learning has gradually replaced traditional methods,has gradually become the mainstream,and is used in all walks of life in society.In recent years,drone inspection has gradually become the most commonly used method for inspection of power grid transmission lines,which greatly reduces the difficulty and risk factor in the inspection process.But for the massive amounts of data captured by drones,what every power grid company urgently needs is to have smarter,faster,and more accurate detection algorithms for key components of transmission lines,so as to be able to cope with complex and heavy detection tasks.This thesis deeply analyzes the characteristics of UAV pictures and the characteristics of key components of transmission lines,such as insulators and anti-vibration hammers,and proposes an effective object detection model for power grid component detection.Main tasks as follows:(1)Due to the confidentiality of the data of various grid companies,there is currently no public data set related to grid transmission lines for researchers to use.In response to this,this thesis constructs a high-quality data set of key components of power grid transmission lines(TLKC).(2)This article deeply analyzes the characteristics of key components in power grid transmission lines.In view of the very large size gap between different types of target objects and the irregular shapes of glass insulators and composite insulators,this thesis proposes a object detection algorithm based on the combination of anchor-based and anchor-free frames(anchor-based and anchor-free fusion module,ABAF).(3)Most of the pictures taken by drones are top views,and the pictures contain a lot of background information.At the same time,the drones will encounter various weathers during high-altitude shooting,which will be affected by it.Aiming at this problem,this thesis proposes a object detection algorithm based on two channels of RGB image and MSR image.Among them,RGB images have good detailed texture information.MSR(Multi-Scale Retinex)is an image enhancement algorithm that can extract high-frequency information in the picture.In this thesis,the features of the two are merged to effectively improve the performance of the algorithm.(4)We use the TLKC data set and PASCAL VOC data set to train the new object detection algorithm proposed in(2)and(3).The experimental results show the superiority of our algorithm compared to other object detection algorithms.