Research And Application Of Semi-supervised Damaged Greenhouse Detection In Transmission Scene

As a critical infrastructure of a nation,the power system plays a vital role in economic and social development.However,one of the challenges during the transmission of electricity is the issue of transmission line safety.Due to extreme weather conditions and human activities,greenhouses are often damaged,causing films to float and hang on transmission lines,posing a serious threat to the safety and stability of the transmission lines.Therefore,to ensure the safe operation of transmission lines,it is essential to detect the damaged greenhouses promptly.Clearly,manual inspection of greenhouses on transmission lines is time-consuming,labor-intensive,and cannot guarantee real-time detection of damages.With the rapid development of computer vision technology,this thesis aims to explore an automated detection model for damaged greenhouses in transmission scenarios,utilizing object detection technology.The current object detection methods can be mainly categorized into two groups:Onestage method and Two-stage method.One-stage method directly detects objects in the entire image by utilizing regression techniques to predict the target position and category.While One-stage method has the advantage of speed,its accuracy is relatively low.On the other hand,Two-stage method extracts candidate regions and then performs classification and regression on these regions.Although Two-stage method has higher accuracy,it is slower in speed.Despite the significant success achieved by these methods in the field of object detection,they all rely on labeled data and belong to fully supervised object detection methods.However,in the task of detecting damaged greenhouses in transmission scenes,labeled data is limited,images with labeled information on damaged greenhouses are relatively scarce.Therefore,directly applying these fully supervised object detection methods may result in overfitting issues and difficult to achieve satisfactory generalization performance.Therefore,this thesis focuses on the recently emerged semi-supervised object detection method,SoftTeacher,to fully utilize unlabeled data in the dataset.The core idea of SoftTeacher is to combine consistency regularization and pseudo-labeling methods in semisupervised learning to train object detection models using a small amount of labeled data and a large amount of unlabeled data.Although SoftTeacher has achieved significant detection results in the field of semi-supervised object detection,it still faces the following three issues:(1)The use of only confidence scores generated by the teacher model as the reliability evaluation standard in unsupervised classification loss calculation lacks comprehensiveness and accuracy.(2)Simple and strong data augmentation methods may result in lack of complexity and diversity in augmented data,making the model only adapt to changes in the training set,leading to overfitting issues.In addition,cropping of key parts of images may occur due to shear or random erasure operations,resulting in the generation of erroneous information and difficulty in ensuring model robustness.(3)The teacher model’s updates are achieved through exponential moving average,resulting in a high similarity between the teacher and student model parameters in the later stages of training,which hinders the effectiveness of consistency regularization methods and limits further improvement of model performance.To address the aforementioned issues,this thesis propose the SoftTeacher+model for semi-supervised greenhouse damage detection.Compared to SoftTeacher,this thesis introduce the unsupervised classification loss function based on joint confidence,the Mosaic augmentation method with KeepAugment,and the differentiated exponential moving average update method.Specifically,this thesis incorporate weighted factors into the unsupervised classification loss function and redefines the standard for reliable evaluation through the joint confidence of the teacher and student models,enhancing the comprehensiveness and accuracy of reliability evaluation.Secondly,this thesis design the Mosaic augmentation method with preservation of relevant information,which retains important information from the original image during data augmentation to reduce the risk of model overfitting and improve the model’s robustness.Additionally,this thesis proposes the differentiated index moving average update method,replacing SoftTeacher’s own exponential moving average method,to reduce the similarity between the teacher and student model parameters,thus further improving the performance of the model.Finally,extensive experiments are conducted on the public dataset MS COCO and the self-built Shelter dataset.The experimental results verify that the proposed methods outperform all baseline methods,demonstrating their practicality and effectiveness.