CG-DINO: Improved Network for Small Object Detection in Uncrewed Aerial Vehicle Imagery Based on DINO

Abstract

Object detection in uncrewed aerial vehicle (UAV) acquired images remains a challenging task due to extreme scale variation, dense object layouts, and complex backgrounds, etc. Although convolutional neural networks (CNNs), transformers, and hybrid model-based detection models have been widely developed and successfully deployed, achieving accurate detection among small or difficult objects in UAV imagery continues to be a pressing task. To address the challenges of small object detection in UAV imagery, we propose CG-DINO, an improved DINO-based detector that introduces a cross-gated transformer block (CGTB). The CGTB couples cross-shaped window self-attention with gated feature refinement, enabling stronger cross-region feature interaction while suppressing background interference. This design enhances feature discrimination for small and densely distributed objects and significantly improves detection accuracy compared with the vanilla DINO framework. In addition, an adaptively reconfigured swin transformer backbone is introduced. In this backbone, the final stage is reformulated with CGTB to more effectively capture the distributional characteristics of UAV imagery. Comprehensive results indicate that our proposed CG-DINO outperforms 13 mainstream detection methods with superior performance across various scenarios. CG-DINO achieves the highest mAP, mAP50, mAP75, mAPS, and mAPM values among the evaluated methods on VisDrone2019-DET. Its mAP50 and mAPS reach 59.3% and 28.3%, respectively, exceeding those of vanilla DINO by 5.1 and 5.0 percentage points. These results further prove the effectiveness of the proposed method in small object detection in UAV imagery.