DCRA-YOLO: DCRA-YOLO: a small object detection network for UAV imagery based on frequency decomposition and deformable attention

In the context of the rapidly evolving low-altitude economy, the capability to detect small objects from unmanned aerial vehicle perspectives has emerged as a critical challenge in the field of computer vision. However, detecting small objects remains a complex task due to limitations such as low image resolution and the susceptibility of targets to being obscured by cluttered backgrounds, which often leads to misidentifications and missed detections. This paper presents DCRA- you only look once (YOLO), an optimized detection framework derived from YOLOv11n, designed to tackle the aforementioned challenges. The model incorporates several novel components: the residual Haar discrete wavelet transform, which enhances fine-grained feature extraction; the deformable large kernel attention, which dynamically improves focus on diverse spatial patterns; and the Conv2Former module, which extends the receptive field from local to global context through convolutional modulation. Additionally, an adaptive threshold focal loss is introduced to emphasize difficult-to-classify samples while maintaining a balanced loss function to improve generalization across complex scenarios. Experimental evaluations on the VisDrone2019 benchmark dataset demonstrate that DCRA-YOLO outperforms YOLOv11n, achieving improvements of 2.9% in precision, 1.7% in recall, and 3.0% in mAP@0.5, respectively. Further evaluations on the tiny-person, SIMD datasets confirm the model’s robustness and generalization capability in small object detection across diverse application scenarios.