TLCFormer: synergizing temporal motion and local contrast for robust infrared video small object detection

Infrared video small object detection remains difficult due to heavy background clutter, low signal-to-clutter ratio (SCR), and severe energy attenuation of 1-4 pixel targets during hierarchical downsampling. Existing frequency-domain filters (e.g., FFT-based Doppler filtering) can be brittle when background textures share similar spectral patterns with target motion. We propose TLCFormer (Temporal-Local-Contrast Transformer), a physics-prior-guided framework that explicitly embeds two key target priors-motion continuity and local intensity extremum-into a transformer detector. TLCFormer introduces: (1) Motion-Aware Difference Attention (MADA), which replaces Doppler filtering with explicit temporal differencing and motion consistency modeling to suppress static clutter while enhancing coherent target motion; (2) a Deep Local Contrast Module (DLCM) that improves SCR by contrasting local maxima against estimated background responses; and (3) a Hybrid Energy-Preserving Mixer that combines max and average pooling during token mixing to mitigate small-target energy dilution. Experiments on RGBT-Tiny and VT-D (Fusion) demonstrate the effectiveness of TLCFormer and its cross-dataset generalization under cluttered infrared scenes.