LDA NET: A lightweight Dual assignment network for extremely small-scale object detection in UAV aerial imagery

Aerial object detection in aerial imagery is a highly complex task due to challenges such as extremely small objects, densely packed distributions, frequent occlusions, and intricate backgrounds. These parameters mostly cause false positives, missed detections, and high computational costs in existing models. To address these issues, particularly for the VisDrone dataset, this work proposes an efficient and accurate LDA Network which is designed to handle extremely small and densely clustered objects. Our approach integrates three key innovations: the SCDown module, the Partial Self-Attention (PSA) mechanism, and an NMS-free detection strategy with dual assignment, which collectively improve both accuracy and efficiency. The SCDown module is a computationally efficient down sampling technique that reduces redundancy while preserving critical spatial details, enabling the model to detect small objects without excessive computational overhead. The PSA mechanism enhances feature extraction by adaptively prioritizing spatially significant regions, thereby improving the detection accuracy of extremely small objects in cluttered and complex visual environments. Finally, the NMS-free detection strategy removes the use of “Non-Maximum Suppression (NMS)” by combining one-to-one and one-to-many predictions, resulting in more precise object localization and faster inference times. Experiments on the VisDrone dataset demonstrate the effectiveness of our model, achieving a “mean average precision” of 54.1%, a “precision” of 78.8%, and a “recall” of 29.8%, outperforming existing YOLO-based models in both accuracy and efficiency. The achieved results indicate the proposed model's ability to address the challenges of aerial object detection, making it a promising solution for real-world.