CDMS-Net: Complementary Detail and Multiscale Selection for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) is difficult because true targets are sparse and weak, whereas cloud edges, sea clutter, building structures, and sensor noise can produce compact target-like responses. This letter proposes CDMS-Net, which combines context-stable and detail-sensitive predictions through a bounded residual adaptive selector using physical saliency, disagreement, and confidence. A fixed calibration provides one low-false-alarm (FA) operating point across datasets. Under a common local protocol, CDMS-Net reduces NUDT-SIRST FA pixels to 338, which is 36.3% and 47.2% lower than MSHNet and DNANet, respectively. Local evaluations using official checkpoints show that WaveTD, DFAwareNet, and HFMNet achieve higher detection accuracy, while CDMS-Net uses 51.2%, 38.6%, and 34.4% fewer parameters, respectively. Parameter sweeps, cue ablations, scene-stratified errors, and complexity measurements characterize this accuracy–compactness tradeoff and its limitations.