GD-PAN: a multiscale fusion architecture applied to object detection in UAV aerial images

Because unmanned aerial vehicle aerial images are captured at high altitude, they contain objects at different viewpoints and in different states, resulting in two major problems in the images: target size imbalance and category semantic imbalance. To alleviate the impact of the above problems on object detection in unmanned aerial vehicle aerial images, this paper integrates the classic path aggregation network with the concept of gather-and-distribute mechanism, proposing a novel feature fusion architecture termed as gather-and-distribute path aggregation network(GD-PAN). GD-PAN leverages the advantages of both fusion approaches, resulting in a more thorough and comprehensive fusion of multi-scale feature maps, dramatically improving the refinement of the network for small target objects. Additionally, we introduce wider coordinate attention to optimize generation of coordinate offset and modulation scalar in deformable ConvNets version 2, which is applied to the basic network module of GD-PAN to enhance the feature extraction and fusion capability of the network for objects in different shapes and states. GD-PAN is a general improvement scheme that can be migrated to object detection algorithms based on the structure of path aggregation networks to improve their fusion ability on multi-scale features as well as small object detection performance. On the VisDrone dataset, the proposed method improves AP by 3.2%, mAP by 2.2%, and AP-small by 2.7% when used for real-time DETR, and when applied to YOLOv7 and YOLOv8, AP-small improves by 2% and 3.3%, respectively.