Research On Image Data Enhancement Technology To Improve The Performance Of Transmission Lines Foreign Objects Detection

Electric energy plays a pivotal role in a country.All aspects of national economic production cannot be separated from electric energy.Transmission lines is the main carriers of power transmission in power systems.Its running status directly affects the operation quality of the power systems.Foreign objects hanging on the transmission lines may cause accidents such as power outages,which will seriously impair the safety and stability of the transmission lines.How to carry out efficient inspection of foreign objects in transmission lines during routine inspection is the key to ensure high-quality operation of power network.Using deep learning technology for foreign objects detection in transmission lines not only has high detection accuracy but also fast detection speed.Using deep learning technology to solve the problem of foreign objects detection in transmission lines not only has high detection accuracy,but also has fast detection speed.However,when the number of training images is limited or insufficient,the detection model may have problems such as low detection accuracy and poor generalization ability.This thesis takes foreign objects in transmission lines inspection as the research object to explore when the number of images in dataset is limited or insufficient.How to use advanced image data enhancement methods to obtain a detection model with higher detection performance.Aiming at the problem of limited or inadequate training data in foreign objects detection model of transmission lines,four data enhancement methods are proposed for training image expansion.The main work of this thesis is as follows:(1)Aiming at the problem of Cutout and Cut Mix may block and replace too many important areas of the original image when the clipping area is large,resulting in the loss of too much important information in the original image,Portion Out and Dynamic Mix methods are proposed.Portion Out improves the defect of removing all pixels in the cropping area in Cutout;Dynamic Mix uses the pixel-level dynamic blending method to associate the cropping region image size with the blending ratio weight.To dynamically change the pixel retention weight of the original image’s clipping region with the size of the clipping region,thereby generating better enhancement samples.(2)Aiming at the Cutout and Cut Mix methods are too random in the selection of clipping regions,and cannot better promote the model to strengthen the learning of the regions with poor discrimination in the image,Saliency Out and Saliency Cut Mix methods are proposed.The purpose of this is to enhance the model’s learning of the less discriminatory area features in the image.These two methods use the saliency detection technology to find part of the saliency area in the image,and then remove or replace the pixels in this area.In this way,the model is guided to enhance the learning of the less discriminant areas in the image.(3)Aiming at the problem of insufficient detection accuracy caused by the limited or insufficient image of the initial transmission lines dataset.Many classical image data enhancement methods and the image data enhancement methods proposed in this thesis are used for image data amplification.Based on this,a data enhancement training set is generated.The total number of images in this training set is 12023,and annotate the target object in the image.A large number of experiments were carried out in the foreign objects y dataset.The results show that using multiple image data enhancement methods can significantly improve the detection performance of multiple mainstream object detection models when the number of images in the training set is insufficient or limited.In the YOLOv5 l object detection model.The model trained using a data augmentation dataset is compared to the model trained using an initial dataset(without using image data augmentation techniques for amplification)m AP@.5 and m AP@.5 :.95 increased by 4.1% and 12.3% respectively.At the same time,a large number of experiments have been carried out in multiple public datasets in the field of image classification.The experimental results show that the four data enhancement methods proposed in this thesis can give the model better robustness and classification performance.