FGADet: Enhanced Fine-Grained Aircraft Detection Network Model with Superior Feature Representation

Fine-grained detection of aircraft is more challenging than traditional object detection tasks, due to the subtle differences among subclasses, significant intra-class variations, and sensitivity to interferences among fine-grained categories. Applying traditional object detection networks, such as YOLO, directly to fine-grained aircraft detection tasks often fails to achieve ideal recognition accuracy, particularly in easily confused subclasses. This paper addresses the distinctive characteristics among aircraft subclasses by designing a feature extraction module named C2CTR to extract more representative features. Firstly, the convolutional modules are improved by introducing the C2CTR with a local attention mechanism, enabling the model to focus on fine-grained regions that distinguish aircraft categories. Secondly, BiFPN connection paths are added to the network's PAN structure, and a channel attention mechanism is utilized, allowing the model's head part to not only comprehensively consider features from different levels but also focus on feature channels that can effectively assist the model in distinguishing aircraft subclasses. Through these designs, the model's representation ability for fine-grained features that distinguish aircraft subclasses is significantly enhanced. Experimental results demonstrate that the proposed FGADet network can effectively identify 20 aircraft subclasses on the military aircraft dataset MAR20, achieving a total mAP of 92.5%. This represents a significant improvement compared to the baseline model YOLOv11x and outperforms the on-the-shelf aircraft object detection networks.