NUDTNet: Physics informed Non-uniform Distinctive Targeting Network for infrared small target detection

Abstract

Infrared small target detection models face a fundamental challenge: their CNN-based backbone applies spatially invariant processing to inherently heterogeneous scenes: targets manifest as concentrated Gaussian-like energy distributions while backgrounds exhibit diffuse patterns from atmospheric scattering. This architectural mismatch causes progressive target feature dilution and degraded localization accuracy in deep networks. We propose Gaussian Recalibration Attention Blocks (GRABs), a physics-informed adaptive convolution framework that transforms static kernels into content-aware operators. GRABs comprise three synergistic components: (1) Feature Decomposition (FD) that separates target and background features through spatially-adaptive bandpass operations, (2) Gaussian Spatial Parameter Predictor (GSPP) that reconstitutes standard convolutions as position-dependent matched filters, and (3) Object Perception Attention (OPA) that provides spatial control signals including probabilistic target presence maps, sub-pixel centroid localization, and anisotropic shape descriptors. We integrate GRABs into the Non-uniform Distinctive Targeting Network (NUDTNet) as the feature extraction backbone, complemented by a Cross-layer Relation Transformer (CLRT) for multi-scale feature fusion. Extensive experiments demonstrate that NUDTNet achieves better performance on NUDT-SIRST, NUAA-SIRST, and IRSTD-1K datasets, validating the effectiveness of incorporating physical imaging principles into neural architectures for infrared small target detection.