HOL-YOLOv11: lightweight small object detection for UAV aerial imagery

With the continuous advancement of science, technology, and UAV aerial photography techniques, the application of object detection in UAV aerial imagery has become increasingly widespread. However, UAV images often present challenges such as small object sizes, high object density, and strong similarity between object features and the background. These factors frequently lead to missed detections and false positives. Moreover, deploying aerial detection models on embedded UAV platforms is often constrained by limited computational resources, resulting in suboptimal accuracy and performance in most existing UAV-based object detection models. To address these challenges, we optimized the YOLO series of models and proposed a lightweight model named HOL-YOLOv11 (Head, Omni-dimensional dynamic convolution, and Linear deformable convolution), specifically designed for small object detection in UAV aerial imagery. First, we designed a tiny-object prediction head and removed the large-object prediction head along with redundant layers, effectively reducing the number of network parameters and accelerating the detection speed. Second, we introduced an omni-dimensional dynamic convolution (ODConv) module into the backbone network to enhance local contextual information while retaining richer semantic information and key visual features, thereby improving small object detection accuracy. Additionally, we integrated a linear deformable convolution (LDConv) module into the neck network to strengthen the fusion of global and local information. Finally, we adopted the Shape-IoU loss function in place of the complete IoU (CIoU) for bounding box regression, further enhancing the localization precision of small objects. To validate the performance of the proposed method, we conducted experiments on the VisDrone2019 and DIOR datasets. Compared to the baseline model YOLOv11s, our improved algorithm achieved increases of 9.6% and 6.9% in mAP@50 on the respective datasets, while reducing the model parameter count by 46.8%.