Research On Remote Sensing Image Object Detection In Complex Scenes

This thesis focuses on the research of high-precision object detection algorithms for remote sensing images in complex scenes.In recent years,with the development of Earth observation technology,object detection in remote sensing images has become the frontier and hotspot of research in this field.Object detection in remote sensing images not only holds significant academic research value within the field of computer vision,but also has wide-ranging applications in military monitoring,national defense construction,intelligent transportation,urban planning,industrial development,public safety,and other fields.Although deep learning-based methods have achieved signifi-cant achievements in this field,performing high-precision object detection in complex remote sensing scenarios,which include a large number of densely arranged,large as-pect ratio,arbitrary-oriented,and blurred objects,remains an important challenge.How to achieve high-precision and efficient object detection in complex remote sensing im-ages has become a current research hotspot.This thesis focuses on object detection in complex remote sensing image scenarios,concentrating on three aspects:tiny object detection,oriented object detection,and semi-supervised oriented object detection,to expand and refine existing remote sensing image object detection algorithms.The main contributions of the thesis include the following aspects:(1)Gaussian Similarity-based Adaptive Dynamic Label Assignment for Tiny Object Detection.Aiming at the problem of traditional object detection networks in remote sensing images exhibiting poor localization quality for tiny objects and an im-balanced distribution of training samples,this thesis proposes a Gaussian Similarity based Adaptive Dynamic Label Assignment for Tiny Object Detection,GS-ADLA.This method initially examines the limitations of the Io U metric in measuring the sim-ilarity of tiny-scale bounding boxes and proposes a Gaussian distribution-based metric to mitigate the impact of minor positional deviations on the localization accuracy of tiny objects.Secondly,an adaptive dynamic label assignment strategy is proposed to alleviate the imbalance of positive and negative samples in the label assignment pro-cess.By integrating these two components,the detection network can obtain sufficient and balanced high-quality training samples,achieving a balance in detection accuracy between objects of different sizes.Experimental results demonstrate that the proposed GS-ADLA method exhibits superior performance on five publicly available aerial re-mote sensing tiny object detection datasets.(2)Frequency-Assisted Dynamic Learning Network for Oriented Object De-tection.Aiming at the problem of detecting oriented objects in remote sensing im-ages and improving the detection accuracy,this paper proposes a Frequency-Assisted Dynamic Learning Network for Oriented Object Detection,FADL-Net.Firstly,by con-sidering the spectral characteristics of remote sensing images,a spatial frequency feature enhancement module is introduced,aiming at extracting a global feature representation with spatial frequency information.Secondly,a geometric adaptive sample mining strat-egy is proposed,which dynamically and adaptively screens high-quality training sam-ples by integrating the object’s geometric information,spatial frequency features,and localization capabilities to improve the accuracy of object localization significantly.Fi-nally,to address the inconsistency between the classification and localization tasks of detectors,a jointly optimized rotational quality loss is introduced.This loss dynamically adjusts the weights of different samples during the training phase,effectively mitigating performance fluctuations and ensuring the reliability of the detector in oriented object detection.Experimental results show that FADL-Net outperforms other methods and achieves state-of-the-art performance.(3)Oriented Object Detection Network based on Semi-supervised Learning.Aiming at the problem of the scarcity of large-scale,precisely annotated data in detect-ing oriented objects in remote sensing imagery,this thesis proposes a Semi-Supervised Oriented Object Detection Network for Remote Sensing Images,S~2O-Det.Existing semi-supervised object detection methods mainly focus on conventional object detec-tion with horizontal bounding box annotations and do not fully consider the arbitrary-oriented objects commonly found in aerial remote sensing imagery.Initially,a task consistency learning strategy is introduced,enhancing the correlation between classi-fication and localization tasks to provide a reliable basis for selecting pseudo-labels.Subsequently,a coarse-to-fine pseudo-label selection strategy is proposed.This method selects specific training pseudo-labels for classification and localization tasks in a divide-and-conquer manner,thereby reducing the impact of noisy labels on network training.Finally,a probabilistic distillation loss is proposed,considering the complex layout of remote sensing images,which fosters consistency in the predictive feature distribution between the teacher and student networks.Experimental results demonstrate that S~2O-Det exhibits outstanding performance under various semi-supervised experimental se-tups with multiple remote sensing image datasets,providing an effective solution for semi-supervised oriented object detection.(4)Consistency-based Semi-Supervised Learning for Oriented Object Detec-tion.Aiming at the problem of ambiguity and unreliability in pseudo-label selection,this thesis proposes a Consistency-based Semi-Supervised Learning for Oriented Ob-ject Detection,CSLO-Det.To further alleviate the misalignment between classifica-tion and localization tasks in semi-supervised oriented object detection,a task align-ment learning strategy is introduced,which includes a feature alignment network and joint optimization loss.This strategy promotes consistency between the two tasks in terms of feature representation and optimization objectives.Secondly,considering the presence of numerous densely arranged,tiny-scaled oriented objects in remote sensing images,a noise-resistant pseudo-label mining strategy is proposed.This strategy em-ploys an adaptive pseudo-label filtering threshold to provide high-quality supervisory information for unsupervised tasks.Furthermore,unlike the traditional semi-supervised learning frameworks that use pseudo-label bounding boxes as supervisory information,CSLO-Det introduces pixel-level dense supervision during the unsupervised training phase.This approach effectively reduces the cumulative bias caused by pseudo-label noise.Experimental results demonstrate that CSLO-Det further enhances detection per-formance in semi-supervised oriented object detection,holding significant theoretical implications and practical value for future research on semi-supervised oriented object detection methods.