OASNet: Orthogonal Attention-Guided Spatial–Semantic Representation Learning Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) plays a critical role in infrared search and tracking (IRST) systems, providing essential support for downstream applications, such as surveillance, early warning, and target tracking. However, most existing methods fail to consider the structural independence between spatial and semantic features, leading to entangled feature representations. Such entanglement limits the network to perform directionally decoupled learning, which is crucial for accurately perceiving small infrared targets from complex backgrounds. To address this limitation, we propose OASNet, an orthogonal attention-guided spatial–semantic representation learning network that constructs the overall architecture as a feature decomposition framework along orthogonal basis directions. The orthogonal attention module captures complementary information along the channel, height, and width dimensions by explicitly modeling their directional independence. Furthermore, the asymmetric orthogonal attention encoding module (AOAEM) and the residual orthogonal attention decoding module (RODM) are introduced to enhance feature hierarchical representation and target localization in an orthogonal decoupled manner. Experiments on NUDT-SIRST, IRSTD-1k, and SIRST demonstrate that OASNet outperforms the state-of-the-art (SOTA) methods in accuracy and robustness.