CLIPAir-YOLO: Semantic-Guided Feature Alignment for Fine-Grained Aircraft Detection in Remote Sensing Images

Abstract

Accurate identification of military aircraft in satellite imagery is critical for defense intelligence and battlefield situational awareness due to the high visual similarity among military aircraft classes. Conventional detectors such as YOLOv8 often fail to distinguish subtle physical details because they rely solely on visual features, which leads to low separability in the feature space. This work proposes CLIPAir-YOLO, a framework that integrates semantic knowledge from a Vision-Language Model (CLIP) into the YOLOv8 detection framework. We introduce the Semantic-Guided Feature Alignment (SGFA) mechanism to align visual features with class semantic prototypes during training. Experimental results on the MAR20 benchmark show that CLIPAir-YOLO improves mAP@0.5 by 0.80% and mAP@0.5:0.95 by 2.08% over the YOLOv8n baseline, with consistent results across three random seeds (std less than 0.001). The method introduces no computational overhead at inference time because the CLIP component is removed after training, making it an efficient solution for military aircraft detection.