A Global Spatial–Temporal Detection Framework for Infrared Small Targets in Complex Ground Scenes

Infrared small-target detection in complex ground scenes still faces significant challenges. Multiframe methods can alleviate the challenges by utilizing temporal motion information when the target is submerged in the background. However, existing multiframe methods have limitations in both extracting temporal motion information and integrating spatial–temporal information, which restricts their detection capabilities. In this article, we propose a global spatial–temporal infrared small-target detection framework. The detection framework consists of two parts: the relative motion pattern extraction (RMPE) module, which computes the relative motion map (RMM) of small targets, and the global spatial–temporal feature fusion network (GST-Net), which takes both the image and RMM as inputs. RMPE calculates dense optical flow maps (OFMs) at each time step and normalizes them, achieving the extraction of small-target relative motion patterns through accumulation. The GST-Net comprises two key modules. The spatial–temporal feature fusion module (STFFM) facilitates the interaction of features at the intermediate stage of feature encoding, while the global spatial–temporal dependency extraction module (GSTDEM) extracts long-range dependencies between targets and backgrounds in deep spatial–temporal semantic feature maps, enhancing the ability to discriminate between the target, noise, and the background. Experiments on multiframe datasets demonstrate that our method can effectively detect small targets under complex backgrounds, outperforming other state-of-the-art (SOTA) methods. The code will be made available at: https://github.com/elvintanhust/GST-Det