DASH-YOLO: Discrete Wavelet Adaptive Suppression and Grouped Adaptive Hyperedges for Infrared Small Target Detection

Infrared small target detection still faces significant challenges under low signal-to-noise ratio and complex backgrounds. First, targets are small and low-contrast; their discriminative information mainly lies in fine-grained high-frequency details, which are easily disturbed by background textures and noise. Second, numerous background nodes in high-resolution shallow features tend to dominate relation aggregation, weakening target-region representation. To address these issues, this article proposes a one-stage infrared small target detection framework, termed DASH-YOLO. Specifically, the proposed method employs a discrete wavelet adaptive suppression strategy to decouple low-frequency structures from high-frequency details, and combines noise estimation to adaptively regulate high-frequency responses. From a physical perspective, this decoupling separates background structural information from local infrared intensity variations, helping preserve small-target edges and grayscale-transition cues while suppressing high-frequency texture noise, thereby enhancing target–background separability. Then, a lightweight high-resolution P2 detection branch is introduced to improve localization and representation for extremely small targets. Furthermore, a grouped adaptive hyperedge modeling mechanism is designed to alleviate background-dominated relation aggregation in dense-node scenarios, enabling more robust high-order feature interaction. Experimental results demonstrate that the proposed method achieves mAP@0.5 scores of 85.62%, 97.50%, and 91.91% on the IRSTD-1 K, SIRST-5 K, and RealScene-ISTD datasets, respectively, while reaching 114.13 FPS with a 640 × 640 input on an RTX 4070 Ti Super, showing a favorable accuracy–efficiency balance.