FWSAM-Net: Wavelet-enhanced SAM2-based framework with frequency-aware adapter for Infrared Small Target Detection

Infrared small target detection (IRSTD) is important for military early warning and civilian monitoring, but remains challenging because targets are small, low contrast, and embedded in complex backgrounds. To address these challenges and mitigate the domain gap arising when visible-light foundation models are transferred to infrared imagery, we propose FWSAM-Net, a wavelet-enhanced SAM2-based framework with a frequency-aware adapter for IRSTD. FWSAM-Net introduces a Frequency-Aware Prompt-Guided Adapter (FPA) that performs Discrete Cosine Transform (DCT)-based channel-spectral decomposition within the SAM2 encoder. It asymmetrically adapts lower- and higher-order coefficients under prompt-guided gating, improving the frozen encoder’s adaptation to and representation of infrared-domain features. To further improve context modeling and multi-scale feature fusion, we propose a Wavelet-based Multi-Scale Context (WMC) module and a Heterogeneous Scale Attention Fusion (HSAF) mechanism. WMC performs multi-scale context modeling through wavelet decomposition and reconstruction, while HSAF facilitates cross-scale feature interaction through heterogeneous attention and channel shuffling. Extensive experiments on three public benchmarks, NUDT-SIRST, IRSTD-1k, and NUAA-SIRST, demonstrate that FWSAM-Net achieves competitive detection performance. It also demonstrates robust cross-domain generalization under cross-dataset transfer and on the progressively degraded Dynamic-ISTD benchmark.