ADD-YOLO: A New Model For Object Detection In Aerial Images

Abstract Currently, object detection models utilized in UAV aerial image tasks encounter challenges such as small and dense objects, as well as interference from complex backgrounds. This paper presents an enhanced model based on YOLOv8s, named ADD-YOLO. In this model, the traditional convolutional layer is replaced with AKConv, which increases adaptability to variations in objects. The C2f_DRAC structure, integrated with AKConv and CBAM, is designed to enhance the model's capability to capture multi-scale contextual information, effectively addressing the issue of background interference. The DABFPN structure incorporates a small target detection layer, which boosts the performance of small object detection and tackles issues related to background interference. Additionally, CIoU-Soft-NMS is introduced to replace the original NMS, enhancing the detection of dense objects and addressing problems such as loss of adjacent prediction boxes and overlap in boundary frame IoU calculations. Extensive ablation studies and comparative experiments were conducted on the VisDrone2019 dataset and the UAVDT benchmark. The results demonstrate that ADD-YOLO outperforms the leading models in UAV aerial image detection tasks, achieving improvements of 15.7% and 7.3% in [email protected] and 13.8% and 5.1% in [email protected]:0.95, respectively, thereby validating the effectiveness of this model.