关键词: 无人机, 深度学习, 红外目标检测, 视频弱小目标检测, 时空特征, 跨层级语义引导

Abstract: Vision-based air-to-air infrared unmanned aerial vehicle (UAV) detection is a key technology for aerial surveillance and swarm confrontation applications. Aiming at the challenges of blurred target boundaries and low target-to-background contrast at a single time instant in infrared video sequences, this paper proposes a spatio-temporal joint feature enhancement framework for air-to-air infrared UAV small target detection method. The proposed approach integrates two dedicated modules into an improved generic U-shaped network architecture. First, a Multi-Level Semantic-guided Spatial Attention Module is designed. High-level semantic information is extracted from down-sampled deep feature maps in the encoder and is used to adaptively enhance the corresponding textured regions of low-level feature maps, producing spatially enhanced features that focus on target regions. Second, a Spatial-Temporally Guided Multi-Head Attention Module is designed. By introducing a temporal window, the spatial-temporally guided multi-head attention module performs cross-attention. It takes the spatially enhanced features of the current frame from the multi-level semantic-guided spatial attention module as the query. It takes high-dimensional decoded features from multiple consecutive frames as the key and value. This design exploits spatial consistency across frames and reinforces the target representation. In this way, the targets’ spatial details and temporal contextual information are jointly enhanced. Extensive experiments are conducted on the Drone subset of the Drone-detection-dataset and the Anti-UAV dataset. The results demonstrate that, compared with state-of-the-art infrared small target detection methods, the proposed method achieves superior performance. Specifically, on the Drone-detection-dataset, our method improves the mIoU by 2.01% and reduces the false alarm rate by 0.72% compared with the strong baseline MFE-Net. On the Anti-UAV dataset, the proposed method achieves a 1.10% improvement in mIoU and a 1.21% reduction in false alarm rate over MFE-Net. These results indicate that the proposed method achieves improved detection accuracy and stability for air-to-air infrared UAV detection.

Key words: UAV, Deep Learning, Infrared Target Detection, Video-based Small Target Detection, Spatio-Temporal Features, Cross-Level Semantic Guidance