Scholarlink

<p>In addressing the challenges of aircraft detection and recognition in remote sensing, particularly where real-world datasets are scarce, we present a novel approach that integrates synthetic aircraft into real satellite imagery through advanced style transfer techniques. While previous research has predominantly focused on terrain, land cover, and weather transfers in satellite imagery, the application of style transfer for generating datasets with distinct objects like aircraft remains under-explored. To bridge this gap, we employ both existing methods and introduce our own adversarially trained Single image Generator-Discriminator framework with multi-scale attention module, SinGAN-MSA, to create satellite images containing synthetic aircraft termed SyntAR images. We rigorously evaluate the effectiveness of these synthetic images by training and testing our framework across various scenarios, including real, synthetic, and hybrid datasets. This comprehensive approach allows us to assess the model’s adaptability across different domains, providing valuable insights into its performance in mixed data environments. Our proposed framework not only enhances the realism and quality of synthetic data through iterative refinement but also tests domain adaptability. Our methodology achieved a 21.31 % improvement in aircraft classification accuracy across five aircraft classes-including four small aircraft classes and one wide-body aircraft class-when using a balanced mix of synthetic and real data, compared to using solely synthetic or limited real data. We observed a 420% increase in the recognition of aircraft classes with limited real data which indicates that proposed framework addresses class imbalance problem. We benchmarked the performance of our framework against state-of-the-art YOLOv8 model in discriminator role. The results demonstrate that our framework is a robust solution, for both data generation and object detection. Results suggest that systematic synthetic data augmentation can improve the performance of aircraft detection and classification in remote sensing images.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SyntAR: Synthetic data generation using multiscale attention generator-discriminator framework (SinGAN-MSA) for improved aircraft recognition in remote sensing images

  • 期刊论文

  • 发表日期：

    2025年4月21日

查看全文
  • Faryal A. Nasir,
  • Khurram Khurshid

摘要

In addressing the challenges of aircraft detection and recognition in remote sensing, particularly where real-world datasets are scarce, we present a novel approach that integrates synthetic aircraft into real satellite imagery through advanced style transfer techniques. While previous research has predominantly focused on terrain, land cover, and weather transfers in satellite imagery, the application of style transfer for generating datasets with distinct objects like aircraft remains under-explored. To bridge this gap, we employ both existing methods and introduce our own adversarially trained Single image Generator-Discriminator framework with multi-scale attention module, SinGAN-MSA, to create satellite images containing synthetic aircraft termed SyntAR images. We rigorously evaluate the effectiveness of these synthetic images by training and testing our framework across various scenarios, including real, synthetic, and hybrid datasets. This comprehensive approach allows us to assess the model’s adaptability across different domains, providing valuable insights into its performance in mixed data environments. Our proposed framework not only enhances the realism and quality of synthetic data through iterative refinement but also tests domain adaptability. Our methodology achieved a 21.31 % improvement in aircraft classification accuracy across five aircraft classes-including four small aircraft classes and one wide-body aircraft class-when using a balanced mix of synthetic and real data, compared to using solely synthetic or limited real data. We observed a 420% increase in the recognition of aircraft classes with limited real data which indicates that proposed framework addresses class imbalance problem. We benchmarked the performance of our framework against state-of-the-art YOLOv8 model in discriminator role. The results demonstrate that our framework is a robust solution, for both data generation and object detection. Results suggest that systematic synthetic data augmentation can improve the performance of aircraft detection and classification in remote sensing images.