IDMSNet: Improved DETR Based on Multi-Scale Feature Fusion for Small Object Detection in UAV Images

Abstract

With the rapid development of UAV technology, UAV image object detection is also facing many challenges, especially for small objects, including dense distribution, feature loss, and background interference. To address these issues, this paper proposes a novel object detection model with dynamic competitive fusion. The framework consists of three major innovations: Firstly, a dynamic competitive multi-scale feature fusion structure is adopted. Combined with a dual-path parallel attention mechanism, it ensures sufficient retention of small object features. Secondly, a small object bias mechanism is introduced. It establishes small object regions through dynamic learning and generates stronger feature responses for small targets. Thirdly, a hybrid matching strategy is introduced to provide the model with more abundant supervision signals and optimize its feature learning capability. Experiments on the VisDrone2019 and COCO2017 datasets demonstrate that, compared with various existing methods, the proposed model improves detection accuracy while maintaining comparable efficiency, and achieves a prominent performance gain especially for small objects. It verifies that the model has good generalization ability, exhibits robustness in handling small targets, and confirms the deployment potential of the proposed model in the practical application of UAV image object detection.