Detection of Vehicles in Remote Sensing Images Using Deep Neural Network

To explore the performance of different YOLO models in detecting vehicles in remote sensing images via deep learning, this study selects Faster-RCNN, YOLOv3, YOLOv5n, YOLOv8n, and YOLOv12n for comparative experiments on the DOTAv1.5 dataset, focusing on three target categories: ship, large-vehicle, and small-vehicle. The models are evaluated for their adaptability to varying scene complexities by analyzing metrics such as precision, recall, and mAP50, alongside computational load and training efficiency. Results indicate that YOLOv8n delivers optimal overall performance with an mAP50 of 0.838, balancing accuracy in simple scenes and anti-interference capability in complex ones, with a computational load of 8.1 GFLOPs. YOLOv12n performs comparably to YOLOv8n in detecting ships and large-vehicles through feature prioritization but underperforms in detecting small-vehicles in complex scenarios. YOLOv3 excels in simple scenes but has high computational demand and long training time, while YOLOv5n’s lightweight design results in insufficient full-scene performance. Faster-RCNN performs the worst. This research provides quantitative references for model selection and optimization in vehicle detection in remote sensing images.