Research And Application Of Semi-Supervised Image Segmentation In Transmission Line Scene

Image segmentation in transmission line scenarios is an important task,which can serve as a preliminary task for safety hazard detection in various transmission line scenarios,such as ice detection of outdoor transmission lines,intrusion detection of trees or engineering vehicles,detection and drawing of electrical components on wires,and disaster monitoring.This article aims to use digital image processing technology and image segmentation technology to study high-precision image segmentation models suitable for industrial application deployment in transmission line scenarios.This study can achieve precise segmentation of wires,towers,and ground in transmission line scenarios,providing support for hidden danger detection in the power system,and ensuring safe and stable operation of the power system.The cost of data annotation for image segmentation tasks is high,and there are annotation errors in some of the annotation data used in this article.Therefore,this article adopts a semi supervised image segmentation method,using fewer annotation data to achieve accurate image segmentation results.This paper mainly introduces the proposed transmission line scene image segmentation method based on semi supervised learning from three aspects:data set preprocessing,image segmentation model and semi supervised image segmentation method.(1)In terms of dataset preprocessing,the original annotated data used in this article is converted from annotated 3D point cloud data.The converted image has the problem of discrete class pixel points,which is not conducive to the training of the image segmentation model and ultimately affects the detection performance of the model.This article adopts the morphological method in digital image processing.Based on the shape characteristics of wires,towers,and ground categories,different preprocessing methods are used for classification,so that in the processed annotated image,each pixel of each category can fully cover its category.At the same time,this article enhances the image of the transmission line scene map to make the characteristics of the wires and towers more obvious,effectively improving the segmentation effect of the wires and towers.(2)In terms of image segmentation models,in order to further improve the effectiveness of the image segmentation model,this article improves the feature extraction network in the image segmentation model by adding an attention mechanism to enable the network to learn the importance of different feature channels.In response to the problem of small pixel widths in the wire category and difficulty in distinguishing image segmentation models,this paper combines the characteristics of the encoder decoder structure with the principle of adaptive spatial feature fusion to fuse more shallow features in the decoder of the image segmentation model.This not only improves the wire segmentation effect,but also improves the overall segmentation accuracy of the image segmentation model.(3)In terms of semi supervised image segmentation methods,this paper analyzes the problems of the self training paradigm in CPS based on the Cross Pseudo Supervision(CPS),which combines the two methods of self training and consistent regularization:in the early stage of model training,false labels appear more errors,misleading the learning of image segmentation models,and proposes a semi supervised image segmentation framework based on random data augmentation,Improved the effectiveness of image segmentation model training.In order to verify the effectiveness of the proposed transmission line scene image segmentation method based on semi supervised learning,this paper conducts a large number of experiments on the real data set of the transmission line scene and uses Mean Intersection over Union(MloU)as the measurement standard.In the experimental section,the effectiveness and necessity of the dataset preprocessing method proposed in this paper were first verified.Through comparative experiments,the superiority of the transmission line scene image segmentation method based on semi supervised learning proposed in this paper is verified.Compared with the best baseline method,the MIoU is improved by about 4%.In the experiment using 1/2 labeled data,the IoU of wires,towers,and ground categories reached 73.26%,77.58%,and 88.12%,respectively.After comparative experiments,a series of ablation experiments were conducted to further demonstrate the effectiveness of the various improvement methods proposed in this paper.