SFERA-Net: Spatial–Frequency Evolutionary Relational Awareness Network for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) remains highly challenging in long-range imaging scenarios due to extremely weak target characteristics and severe background interference. Although recent deep learning-based methods have achieved remarkable progress through hierarchical representation learning, subtle target cues are still susceptible to progressive attenuation and interference contamination during feature extraction, leading to insufficient target–background discrimination under complex and nonstationary scenes. To address these challenges, we propose a spatial–frequency evolutionary relational awareness network (SFERA-Net), which improves small target representation through coordinated frequency-domain analysis and spatial contextual dependency modeling. Specifically, within the downsampling stage, a collaborative wavelet-attentive downsampling (CWAD) module is introduced to adaptively preserve subtle high-frequency target cues during feature compression. This is followed by a focal evolution awareness block (FEAB) as the feature extraction backbone, which enhances local salient responses using a multiview selection mechanism while modeling long-range contextual evolution through hidden-state dynamics. After feature fusion, a grouped relational affinity module (G-RAM) is employed to refine multiscale representations by constructing an affinity matrix, explicitly modeling structural relationships and improving target–background separability. Extensive evaluations on the IRSTD-1k, NUAA-SIRST, and NUDT-SIRST datasets demonstrate that SFERA-Net consistently outperforms contemporary state-of-the-art (SOTA) methods while maintaining competitive computational efficiency.