Bidirectional Motion Perception Transformer Network for Multi-Frame Infrared Small Target Detection

Multi-frame infrared small target detection (MIRSTD) aims to detect weak and fast-moving targets by leveraging temporal information across consecutive frames. However, existing MIRSTD methods primarily focus on intra-frame spatial contrast while insufficiently modeling temporal dependencies, leading to weak motion perception and poor temporal coherence. To address these issues, we propose BMPTransNet, a Bidirectional Motion Perception Transformer Network that effectively captures motion information across consecutive frames. First, a multi-scale encoder with a Spatial-embedded Channel Interaction Block (SCIB) is employed to generate fused spatial features, which are redistributed through a pyramid structure to improve the semantic distinction. Second, a Bidirectional Motion Perception Branch (BMPB) is introduced to align multi-frame features, extracting salient temporal cues while suppressing dynamic background interference. Third, a multi-frame joint iterative decoder augments initial queries with additional motion-guided queries derived from shallow features and temporal motion information. Finally, a self-supervised motion-consistency learning strategy guides the network to maintain temporal coherence and spatial accuracy across consecutive frames. Experimental results on three public datasets demonstrate that BMPTransNet achieves accurate detection of moving infrared small targets.