E-FPN: an enhanced feature pyramid network for UAV scenarios detection

Unmanned aerial vehicle (UAV) is versatile machines that capture aerial images with a bird’s eye view of objects from various perspectives and heights. They are widely utilized in both military and civilian applications. As the domain of computer vision advances, object detection becomes a solid mainstream for UAV applications. However, due to the flight altitude of the drone and the variation of shooting angle, drone images often contain small size, dense, and confusing targets, resulting in low detection accuracy. In this article, we propose a new approach called enhanced feature pyramid network (E-FPN) for detecting objects in UAV scenarios. Our E-FPN architecture incorporates the Simplified Spatial Pyramid Pooling-Fast (SimSPPF) structure into the backbone, enabling the extraction of features at four different scales. These features are subsequently passed through different layers of the E-FPN neck, facilitating the interaction between shallow and deep features. This process gains four distinct feature representations. Firstly, the input images are pre-processed by data augmentation. Then, the CSPDarknet53 with SimSPPF is used as the backbone to extract the multi-scale features from the visuals. Secondly, the integration of Cross-Stage Partial Stage modules into the E-FPN framework enhances the network’s ability to capture target details. The E-FPN neck’s top–down pathway integrates features from diverse layers and scales of the backbone, generating three richer and multi-scale representations of intermediate features. Meanwhile, its bottom–up pathway fuses semantic information from various layers and scales into four feature maps of equal channels but differing scales. Finally, a detector head is added to improve the accuracy of the detection, and the final results are obtained. The experimental results demonstrate that E-FPN-N achieves mAP50-95 performance of 37.7% on the MS COCO2017 dataset. Moreover, the precision and mAP50 of our model on the VisDrone validation dataset reached 67.5% and 62.0%, respectively.