| Unmanned aerial vehicle(UAV)remote sensing image-based object extraction has been widely applied in various fields,and deep learning methods are promoting automation and intelligence development in multiple industries.How to apply deep learning methods to achieve intelligent object extraction has become an urgent scientific problem in the process of automating UAV remote sensing image interpretation.However,UAV remote sensing images generally suffer from the issue of low-quality samples,and deep learning models trained on low-quality samples cannot accurately perform object extraction.This paper proposes a method based on semi-supervised learning by training a model with a few high-quality samples and many unlabeled samples.For object extraction research,this study employs typical UAV remote sensing images obtained from rural land consolidation quality inspection data in Hechi,Guigang,and Guiping cities.Specifically,this paper:(1)Constructs a UAV remote sensing image object extraction dataset for deep learning,including fully supervised and semi-supervised datasets;(2)Integrates pseudo-label generation methods and consistency regularization methods to design a semi-supervised UAV remote sensing object extraction model;(3)Implements the semi-supervised object extraction model for object extraction and validation.The main conclusions of this study are as follows:(1)Based on the rural land consolidation quality inspection data images and the poor overlaying of object vector patches,this paper constructed a UAV remote sensing image object extraction deep learning dataset.The fully supervised dataset contains1,006 image samples with a noise rate of 69%.The semi-supervised dataset includes440 labeled images and 3,789 unlabeled images,totaling 4,229 images.The object categories in both datasets are evenly distributed,and the study provides sample data for training deep learning models for other UAV remote sensing image-based object research.(2)Combining the pseudo-label generation method and consistency regularization method in semi-supervised learning,this paper designed a semi-supervised UAV remote sensing image object extraction model considering the easy confusion of landscape features and the diverse complexity of object features in images.This paper proposed a high-quality pseudo-label generation algorithm that combines "strong data augmentation" consistency regularization and a "multi-epoch comparison calculation,from easy to difficult" high-quality pseudo-sample selection algorithm to improve the quality of selected samples,thereby enhancing the performance of the object extraction model.(3)Intelligent extraction of objects in UAV remote sensing images.This paper used DeepLabv3+ as the semantic segmentation network to implement the semisupervised learning-based object extraction model.In the result validation,compared with the best-performing fully supervised learning model,the semi-supervised learning model designed in this study improved mIoU by 19.1%,reaching 66.2%.In object extraction results,the models highest precision reached 0.803,the highest recall reached 0.770,and the F1 score reached 0.708.This paper conducts UAV remote sensing image object extraction based on semisupervised deep learning methods.In the pseudo-label generation method,this study introduced consistency regularization methods to improve the quality of pseudo-labels,thereby enhancing object extraction accuracy.This research can provide references and ideas for future related studies. |
