AMFNet: adaptive multi-scale fusion network for small object detection

In Unmanned Aerial Vehicle (UAV) aerial and remote sensing images with complex backgrounds, small object detection faces challenges such as local feature ambiguity, multi-scale semantic mismatch, and weak adaptability to geometric deformation. Therefore, we propose an Adaptive Multi-Scale Fusion Network ( AMFNet ), which enhances feature representation of small objects from three aspects: feature enhancement, cross-scale fusion, and deformation modeling. Firstly, the Spatial Structure Enhancement Module ( SSEM ) captures global dependencies using single-head self-attention, while heterogeneous convolution extracts local details, thereby achieving complementary global–local feature modeling. Secondly, we construct a Dual-guided Cross-scale Enhancement Path ( DCEP ), which aligns and complements multi-scale features through deep semantic guidance and shallow structural feedback, together with a cascaded loop fusion strategy. Finally, an A2-Dynamic Adaptive Module ( A2-DyAM ) is introduced, adapting to geometric deformation using deformable convolution, and a Position Adaptive Transformation (PAT) mechanism refines the features to further improve geometric accuracy. Experiments on the VisDrone2019 dataset demonstrate that AMFNet improves mAP50 by 12.0% over the baseline YOLOv12 and 12.7% over YOLOv26. These results indicate that the proposed method maintains stable detection performance under low-discriminative features and geometric variations.