Spatial–Frequency Inductive Bias-Guided Cross-Domain Representation Learning for Infrared Small Object Detection

Infrared small object detection (ISOD) plays a crucial role in military reconnaissance, security surveillance, and remote sensing monitoring, where weak thermal responses and complex backgrounds impose significant challenges. The recent self-supervised vision foundation model DINOv3 has demonstrated remarkable generalization ability across various visual tasks. However, directly transferring it to ISOD still remains challenging due to substantial cross-domain discrepancy between visible and infrared imagery, as well as the limited granularity of foundation features in capturing subtle thermal variations. To address these issues, this study proposes a spatial–frequency inductive bias-guided network (SFI-Net) based on DINOv3 for cross-domain representation learning in infrared small object detection. Instead of conventional domain adaptation strategies, SFI-Net explicitly models infrared-specific inductive biases in both spatial and frequency domains to enhance transferred representations. First, a spatio-frequency hybrid adapter (SFHA) is designed and embedded across multiple layers of the frozen backbone to learn infrared-specific inductive biases within distinct subspaces. Second, a feature compensation strategy with an auxiliary convolutional branch is devised to compensate for the limitation of DINOv3 in capturing multi-scale fine-grained features. Extensive experiments on the IRSTD-1K and NUDT-SIRST datasets demonstrate that the proposed SFI-Net outperforms state-of-the-art methods in both detection accuracy and computational efficiency while exhibiting strong cross-scenario generalization capability.