.png){kind=icon}
.png){kind=icon}
Return
ST-PINet: Spatiotemporal Physics-Informed Network for Moving Infrared Small Target Detection via Endogenous Decoupling
.png){kind=icon}
PRE
AI
J
Z
Z
G
H
DOI：10.1109/tgrs.2026.3716921.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Moving infrared small target detection (MIRSTD) remains highly challenging due to the extremely small target size and the ultralow contrast against cluttered backgrounds. Existing deep learning approaches are predominantly treated as purely image-driven pipelines, where physical modeling components are introduced merely as isolated “plug-in” modules, thereby overlooking the intrinsic coupling between radiative transfer and sensor imaging processes. To this end, we propose the spatiotemporal physics-informed network (ST-PINet). By incorporating physical degradation inversion modeling and physics-constrained decoupling, ST-PINet endogenously embeds the infrared imaging mechanism into the feature learning process, thereby restricting the network optimization within a physically feasible solution space. Specifically, we design a thermodynamic endogenous convolution (TEC) that parameterizes the point spread function (PSF) and atmospheric degradation as learnable differential physical kernels. Furthermore, a spatial adaptive noise gating (SANG) mechanism is introduced to simulate the nonuniform response characteristics of infrared detectors. In addition, we develop a physics-driven triple decoupling (PTD) strategy based on the law of energy conservation, achieving pixelwise endogenous separation of targets, background, and noise. Extensive experiments on multiple benchmark datasets and under complex, extreme conditions demonstrate that the proposed method achieves superior detection performance, strong robustness, and enhanced physical interpretability.
Keywords:
Convolutional neural networks (CNNs)
moving infrared small target detection (MIRSTD)
physics-informed neural networks (PINNs)
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W
Organization
Cited Papers