GeoRankNet: geometry-aware feature ranking network for infrared small-target detection

Infrared small target detection remains a persistent challenge due to inherently low signal-to-clutter ratios, limited spatial resolution, and complex background interference. To address these issues, we propose GeoRankNet, a Geometry-Aware Feature Ranking Network that integrates three novel components: (1) Geometry-Aware Deformable Difference Convolution (GAD2C), which learns adaptive sampling offsets to align convolutional kernels with irregular target contours, thereby mitigating grid artifacts and enhancing localization precision; (2) Dynamic Feedback-Driven Top-K Attention, a training-aware feature ranking mechanism that adjusts the pooling cardinality in real time based on IoU and loss feedback, balancing contextual richness and noise suppression across training stages; and (3) Gated Multi-Attention Fusion (GMAF), which jointly leverages channel, spatial, and global context attention and uses a lightweight gating network on globally pooled features to adaptively fuse multi-scale encoder and decoder representations, amplifying true target cues and suppressing false alarms. Extensive experiments on multiple public single-frame infrared small target (SIRST) benchmarks demonstrate that GeoRankNet consistently outperforms previous methods in both detection accuracy and robustness under challenging infrared small-target scenarios.