A UAV Image Object Detection Algorithm Based on Deep Diverse Branch Block and Multi-Scale Auxiliary Feature

Unmanned Aerial Vehicle (UAV) image object detection has been widely applied in many fields. However, compared with ordinary natural images, UAV images often exhibit complex backgrounds, a predominance of small objects, and significant variations in target scales, which cause traditional detection algorithms to easily suffer from missed or false detections with insufficient accuracy. To address these issues, this paper proposes a novel UAV image object detection algorithm named DMA-YOLO based on the YOLOv8s model, incorporating a deep diverse branch block and multi-scale auxiliary feature. First, a DF-C2f module integrating a deep diverse branch block and an adaptive fine-grained attention mechanism is designed to enhance small object detailed feature extraction. Second, a multi-scale auxiliary feature pyramid network (MAFPN) reconstructs the neck structure to strengthen multi-scale feature fusion and interaction, mitigating the impact of target scale variations. Finally, a dynamic detection head (DyHead) optimizes detection performance and model robustness, and the EIoU loss replaces the original CIoU to enhance bounding box regression accuracy and stability. Ablation experiments on the public VisDrone2019 dataset show that DMA-YOLO achieves a 3.7% increase in mAP50 and 2.8% in mAP50:95 compared with the baseline, with negligible changes in parameter counts and computational complexity. Comparative experiments with mainstream detection models and UAV-specific state-of-the-art algorithms confirm DMA-YOLO’s superior detection accuracy. Further experiments on the RSOD dataset validate its generalization capability and stability across diverse data distributions, highlighting its applicability in complex UAV object detection scenarios.