YOLO-RLDW: An Algorithm for Object Detection in Aerial Images Under Complex Backgrounds

Aiming at the challenges of low detection accuracy, susceptibility to complex background interference, difficulty in detecting small objects, and multi-scale object issues in aerial images, our proposed an improved YOLOv8-based object detection algorithm, named YOLO-RLDW. Leveraging the advantages of Receptive Field Attention Convolution (RFAConv), we introduce a feature extraction module called C2f-RFA to enhance the capability of extracting features of small objects in aerial images. Inspired by the idea of Large Sparable Kernel Attention (LSKA), we design the SPPF-LSKA module with long-range dependency and adaptive capability, which effectively mitigates the interference of aerial backgrounds on object detection. The detection head of YOLOv8 is replaced with a Dynamic Head, and different detection heads are unified using scale, spatial, and task attention mechanisms, dynamically adjusting their detection strategies according to the sizes of targets, thereby further improving the model’s generalization and adaptability. Finally, we adopt the Wise-IoU loss based on a dynamic focusing mechanism for bounding box regression, which accelerates the convergence speed of the model while improving the localization capability of multi-scale objects. Experimental results demonstrate that on the VisDrone2021 dataset, the proposed algorithm achieves improvements of 5.5%, 3.9%, 5.4%, and 3.7% in precision (P), recall (R), mean average precision (mAP50), and mAP95, respectively, compared to the original algorithm. On our self-built remote sensing image dataset RSI, the accuracy, recall, and mean average precision reach 94.2%, 91.0%, and 95.4%, respectively, demonstrating good performance in detecting objects in aerial images. Comparison with other mainstream object detection algorithms validates the effectiveness and superiority of the proposed method.