@article{YIN2026, 
author = {Xingyu YIN and Ren JIN and Zhaochen CHU and Mingdong JIA and Xinfu LIU},
title = {Spatio-temporal feature joint enhancement for air-to-air infrared UAV detection},
year = {2026},
journal = {Acta Aeronautica et Astronautica Sinica},
volume = {47},
number = {20},
keywords = {UAV, deep learning, infrared target detection, infrared video small target detection, spatio-temporal features, attention mechanism},
url = {https://www.sciopen.com/article/10.7527/S1000-6893.2026.33342},
doi = {10.7527/S1000-6893.2026.33342},
abstract = {Vision-based air-to-air infrared Unmanned Aerial Vehicle (UAV) detection is a key technology for aerial surveillance and swarm confrontation. 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 feature joint enhancement method for air-to-air infrared UAV target detection. The proposed method integrates two modules into an improved generic U-shaped network architecture. First, a Cross-Level Semantic-Guided Spatial Attention (CLSA) module is designed. During the encoder stage, semantic information from deep feature maps generated through downsampling is used to adaptively enhance the texture features in the corresponding regions of shallower feature maps, yielding spatially enhanced feature maps and focusing the features on target regions. Second, a Spatial-Temporally Guided Multi-Head Attention (ST-MHA) module is designed. During the decoder stage, a temporal window is introduced, and cross-attention is performed between the high-dimensional decoded features of consecutive frames within the temporal window and the spatially enhanced feature maps of the current frame output by the CLSA module. Multi-frame spatially consistent features are used to enhance the current target representation, thereby achieving joint feature enhancement of spatial details and temporal context. Extensive experiments are conducted on the UAV subset of the Drone-detection-dataset and the Anti-UAV dataset to compare the proposed method with state-of-the-art infrared small target detection methods. The results demonstrate that the proposed method improves the mIoU by 2.01% and reduces the false alarm rate by 0.72×10-4 compared with the competitive MFE-Net on the UAV subset of the Drone-detection-dataset. On the Anti-UAV dataset, the proposed method achieves a 1.10% improvement in mIoU and a 1.15×10-4 reduction in false alarm rate over MFE-Net. These results indicate that the proposed method effectively improves the detection accuracy and stability of air-to-air infrared UAV target detection.}
}