An efficient YOLOv12-based framework for detecting extremely small-scale objects

Abstract

Abstract Object detection in aerial imagery, particularly from UAVs and drones, presents unique challenges such as small and extremely small-scale object distributions. To address these drawbacks, the present research proposes an efficient YOLOv12 model. The first core contribution of our work is the introduction of the Area-Attention C2f (A2C2F) module, which fuses multi-head MLP blocks with localized area-attention mechanisms to strengthen spatial feature learning while maintaining lightweight computation. Second, the model integrates the Cross Stage Partial with Kernel size 2 (C3K2) module, a lightweight variant of the C3 block, which reduces convolutional complexity and computational cost while maintaining feature extraction capability for efficient detection. Third, a multi-scale fusion strategy by stacking A2C2F blocks along with Concat and Upsample operations ensures the preservation of high-resolution features for detecting small-scale objects, a common challenge in drone-captured data. Fourth, a novel detection head with decoupled classification and regression branches, enhanced by attention-guided feature fusion, improves prediction consistency and boosts inference speed. Extensive experiments conducted on the VisDrone dataset demonstrate that the proposed framework effectively detects objects with dimensions smaller than 3–5 pixels, where traditional YOLO variants typically fail, proving its robustness in extremely small-scale detection. The proposed model achieves a Precision of 69.1%, Recall of 48.5%, and F1-score of 56.99%, outperforming recent models. Notably, it achieves a mAP@50 of 58.8% and mAP@0.5:0.95 of 40.9%, showing substantial improvement over prior methods. Despite utilizing 59.1 M parameters and 198.6 GFLOPs, the model maintains an efficient inference rate of approximately 40 FPS on the A100 GPU, confirming its suitability for real-time UAV edge deployment. When compared to two-stage and anchor-based models, our model delivers higher AP scores at a lower input resolution of 640 × 640 pixels, making it a highly effective and computationally efficient framework for UAV-based small and extremely small object detection.