KL-YOLO: A Lightweight Adaptive Global Feature Enhancement Network for Small-Object Detection in Low-Altitude Remote Sensing Imagery

Using advanced object detection technology, drones can achieve long-distance, efficient monitoring and target recognition, which widely applied in fields such as military, agriculture, and rescue works. However, object detection based on UAVs is confronted with a multitude of challenges, such as motion blur due to high speed, high proportions of small targets, difficulty in feature extraction, and decreased recognition accuracy due to object occlusion. To address these challenges, this paper proposes a lightweight adaptive global feature enhancement network (KL-YOLO) based on YOLOv8n for small object detection in low-altitude remote sensing images.The proposed model adds a Tiny Object Detection Layer (TODL) to improve the ability to collect and detect small target semantic information by extracting more fine-grained information from the feature maps. A Global Attention Mechanism (GAM) is introduced in the backbone to suppress irrelevant background information and highlight target regions, enhancing adaptability to varying scales. In addition, the WIoUv3 bounding box regression (BBR) loss function is introduced to utilize its weighting mechanism, dynamically adjusting the loss contributions of targets of different sizes, in order to improve the model’s sensitivity to small objects during the regression stage. Finally, a Dynamic Head (DyHead) integrating self-attention is employed to adaptively aggregate features at different scales and generate weights, addressing issues related to feature map resolution loss due to downsampling. Extensive experiments conducted on the VisDrone2019 dataset demonstrate that the proposed KL-YOLO model achieves an mAP@0.5 of 43.1% with only 3.16MB parameters, representing an 8.0% improvement over the baseline. Furthermore, the model improves precision by 9.7% and recall by 5.3%, effectively enhancing detection performance for small objects.