SFJR-Net: Spatial–Frequency Joint Representation Network for Infrared Small Target Detection

Abstract

Accurate infrared small target detection (IRSTD) is critical for all-weather perception systems. However, conventional deep networks tend to exhibit low-pass filtering behavior, leading to irreversible attenuation of fragile signatures in small targets. Meanwhile, complex clutter overwhelms discriminative cues, submerging faint targets into the background. To address this challenge, we propose the spatial–frequency joint representation network (SFJR-Net). Specifically, we introduce the spatial–frequency preserving downsampler (SFPD) for high-fidelity feature decimation. By combining wavelet-based spectral decomposition with topology-preserving spatial folding, SFPD suppresses aliasing and preserves fine-grained structural details. Subsequently, the context-singularity decoupling mechanism (CSDM) is employed to dynamically separate target signals from background clutter. Guided by a spatial arbitration map, this mechanism utilizes asymmetric receptive fields (RFs) to prevent localized target features from being assimilated by global background semantics. Finally, the spectral-gated reconstruction fusion (SGRF) reconstructs high-frequency energy maps by exploiting wavelet invertibility. Acting as an adaptive spectral gate, SGRF restores sharp structural details, thereby enabling spatial–frequency refinement. Extensive experiments on SIRST, NUDT-SIRST, and IRSTD-1k benchmarks demonstrate that SFJR-Net achieves state-of-the-art performance for precise small target localization in complex thermal scenes.