ATG-Net: Improved Feature Pyramid Network for Aerial Object Detection

Object detection in aerial images is gradually gaining wide attention and application. However, given the prevalence of numerous small objects in the Unmanned Aerial Vehicle (UAV) aerial images, the extraction of superior fusion features is critical for the detection of small objects. However, feature fusion in many detectors does not fully consider the specific characteristics of the detection task. To obtain suitable features for the detection task, the paper proposes an improved Feature Pyramid Network (FPN) named ATG-Net, which aims to improve the feature fusion capability. Firstly, we propose an Adaptive Tri-Layer Weighting (ATW) module that adaptively assigns weights to each layer of the feature map according to its size and content complexity. Secondly, a Triple Feature Encoding (TFE) module is implemented, which can fuse feature maps from three different scales. Finally, the paper incorporates the Global Attention Mechanism (GAM) into the network, which includes improved channel attention mechanisms and spatial attention mechanisms. The experiments are conducted on the VisDrone2020 dataset, and the result shows that the network significantly outperforms the baseline detector and a variety of popular object detectors, which significantly improves the feature fusion capability of the network and the detection accuracy of small objects.

@article{Zheng2024,
  title     = {ATG-Net: Improved Feature Pyramid Network for Aerial Object Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {11},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Junbao Zheng and ChangHui Yang and Jiangsheng Gui},
  doi       = {10.14569/IJACSA.2024.01511116},
  url       = {https://doi.org/10.14569/IJACSA.2024.01511116}
}