STDF-Net: Spatio-Temporal Decoupling and Frequency-Aware Fusion Network for Multiframe Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) plays an important role in a wide range of civilian and military applications. However, the accurate detection remains challenging because infrared (IR) targets usually occupy only a few pixels and exhibit weak contrast, while complex background structures can generate target-like responses. Existing single-frame methods process each image independently and cannot make full use of temporal information across consecutive frames, often resulting in missed detections and false alarms. To address these issues, we propose a spatio-temporal decoupling and frequency-aware fusion network (STDF-Net) for multiframe IR small target detection. First, a spatio-temporal decoupled gated convolution (STDG-Conv) utilizes directional convolutions along the spatial and temporal dimensions to capture complementary spatial structures and interframe correlation. The resulting directional features are jointly regulated by a shared gating mechanism (SGM) and fused to exploit temporally consistent information across consecutive frames, thereby reducing missed detections. Then, a temporal weighted aggregation (TWA) module assigns learnable weights to different temporal positions and aggregates the corresponding features, producing a unified representation for the subsequent encoder. Next, within the encoder, a frequency-spatial self-attention fusion block (FSSFB) decomposes the input features into low- and high-frequency components using the Haar wavelet transform (HWT). These high-frequency components are fused and then refined by our presented high-frequency preserving self-attention (HFPSA) mechanism to guide spatial feature interaction. The resulting spatial features are subsequently combined with the low-frequency component to improve target–background discrimination and reduce false alarms. Finally, the decoder transforms the encoded features into a pixelwise probability map, which is binarized to obtain the final binary segmentation map. Experimental results on two benchmark multiframe IR small target datasets, TSIRMT and NUDT-MIRSDT, demonstrate that STDF-Net achieves better overall detection performance than the compared state-of-the-art methods, with <inline-formula> <tex-math notation="LaTeX">$F_{1}$ </tex-math></inline-formula>-scores/IoUs of 0.741/0.588 and 0.905/0.826, respectively. The corresponding <inline-formula> <tex-math notation="LaTeX">$P_{d}$ </tex-math></inline-formula> values are 0.774 and 0.985, while the <inline-formula> <tex-math notation="LaTeX">$F_{a}$ </tex-math></inline-formula> values are <inline-formula> <tex-math notation="LaTeX">$3.475~\times ~10^{-4}$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$2.715~\times ~10^{-6}$ </tex-math></inline-formula>, respectively. The source code is available at: <uri>https://github.com/MinjieWan/STDF-Net</uri>