AGI-NeXt: A Next Generation Attention Enhanced Gated Inception Network for Air-to-Air UAV Detection
Abstract
The rapid growth of Autonomous aerial vehicles (AAVs) in civilian and defense domains has intensified the need for reliable detection systems to ensure airspace security and situational awareness. Detection in air-to-air scenarios remains highly challenging due to scale variations, rapid motion, and background clutter. To address these challenges, this work proposes AGI-NeXt, a next-generation deep learning architecture that combines dual feature extraction, adaptive channel retention, and a reparameterized detection head. The dual extraction block enhances semantic and spatial representations, while the GatedCaSE module selectively retains gradient-rich channels, and the CoGA module preserves directional context with minimal overhead. Together with the IConvX fusion block, AGI-NeXt achieves robust multiscale feature learning. Evaluations on DetFly, LRDDv1, and UAVFly datasets are conducted using metrics, such as precision, recall, mAP, accuracy, inference speed, and model size. Results show that AGI-NeXt surpasses YOLO-based and state-of-the-art methods, delivering superior accuracy, strong generalization and robust performance across diverse UAV sizes, motion patterns, and complex environments. Ablation studies show that AGI-NeXt balances efficiency and accuracy via adaptive channel retention, enabling efficient UAV detection.