A dual-stream spatio-temporal cross attention network for moving infrared dim and small target detection

Abstract

Infrared small target detection (IRSTD) is a critical technology in modern early warning and sensing systems. Existing single-frame detection methods rely solely on spatial domain information and struggle to effectively distinguish targets from background clutter in complex scenarios. While multi-frame detection methods introduce temporal cues to enhance detection performance, they still exhibit limitations in spatio-temporal feature extraction and fusion strategies. To overcome these challenges, we propose a novel Dual-Stream Spatio-Temporal Cross Attention Network (DSCA-Net). Specifically, we first design a Dual-Stream Feature Extractor (DSFE) that employs parallel 3D and 2D convolutional networks to extract temporal motion information and spatial detail features, respectively, effectively achieving decoupled representation of spatio-temporal features. Second, we introduce a Temporal Attention (TA) module that employs an adaptive weighting mechanism to dynamically enhance motion responses at critical moments, significantly improving the network’s perception capability for target motion characteristics. Finally, we construct a Spatio-Temporal Cross-Attention Fusion (STCAF) module that achieves deep interactive fusion between static spatial features and dynamic motion features through standard cross-attention and position-guided cross-attention mechanisms. Systematic evaluation on three public datasets demonstrates that the proposed DSCA-Net achieves competitive performance compared to mainstream methods, with particularly notable improvements on the IRDST dataset, validating its effectiveness in complex scenes.