ADFPN-YOLOv8s: Small Object Detection From Low-Altitude UAV Perspective

Abstract

To address the issues of easy loss of small object information during feature fusion, severe complex background interference, and insufficient small object features in low-altitude UAV imagery, an improved YOLOv8s model named ADFPN-YOLOv8s is proposed for small object detection. An Attention-based Dynamic Feature Pyramid Network (ADFPN) is designed, which constructs a feature enhancement path guided by a Simple, Parameter-Free Attention Module (SimAM). Through a dynamic weighting mechanism, selective cross-level feature fusion is realized, and the Leaky ReLU activation function is applied to impose nonlinear constraints on the fusion weights, thereby enhancing the feature response of small objects and reducing information loss during hierarchical transmission. Meanwhile, a Multi-scale Attention Spatial Pyramid Pooling (MASP) fusion module is developed by means of feature recalibration and attention fusion. Efficient Multi-Scale Attention (EMA) is embedded into a small-scale spatial pyramid pooling framework to suppress complex background noise while preserving the detailed contextual information of small objects. Furthermore, a high-resolution P2 detection layer is introduced to fully exploit low-level fine-grained pixel information, compensate for the insufficiency of small object features, and improve the model’s perception capability for small objects. Experimental results on the VisDrone2019-DET dataset show that, compared with the baseline YOLOv8s, the precision, recall, mAP@0.5, and mAP@0.5:0.95 of ADFPN-YOLOv8s are improved by 4.2%, 2.7%, 3.5%, and 1.4%, respectively. On the HIT-UAV dataset, the mAP@0.5 reaches 95.6%; on the SeaDroneSee dataset, the mAP@0.5 reaches 79.7%. The results on multiple datasets demonstrate that the proposed model achieves favorable detection accuracy and generalization ability for small object detection from the UAV perspective.