Knowledge-Enhanced RT-DETR for Infrared Small Target Detection Driven by Multilevel Domain Priors

Abstract

Infrared small target detection (IRSTD) remains an arduous task characterized by diminished target signatures, suppressed signal-to-noise ratios, and intricate environmental clutter. Conventional data-driven paradigms often fail to resolve dim targets from overwhelming background interference, as the weak signals are easily submerged by noise. In this study, we present a knowledge-enhanced RT-DETR (KE-RT-DETR) framework that synergistically integrates infrared physical priors into a multilevel neural architecture, facilitating robust target-background separation in low-SNR scenarios. At the input stage, a statistical-physical context enhancement (SPCE) module is employed to augment target saliency by capturing directional responses and statistical distributions. This is complemented by a soft-gated SCConv (SG-SCConv) during feature extraction, which effectively decouples high-frequency target components from low-frequency background fluctuations to strengthen feature discriminability. Furthermore, a gradient-adaptive reparameterized central difference convolution (GA-RepCDC) is introduced in the feature fusion layer to accentuate local contrast through differential modeling while maintaining high inference throughput via structural reparameterization. The experimental results on the DAUB and IRSTD-1k datasets demonstrate that the proposed method achieves superior performance against the compared methods. The code and dataset will be made publicly available at https://github.com/Jixiang975231/KE-RT-DETR