DMS-OGSTV: A Novel Infrared Small Target Detection Based on Spatial-Temporal Tensor Model

Detecting small targets in complex infrared backgrounds is challenging due to edge aliasing and noise interference. Tensor decomposition methods show potential but have limitations in these conditions. This paper proposes a dynamic motion saliency infrared tensor model that integrates temporal information to address these challenges. The method formulates target detection as a low-rank sparse tensor decomposition problem in the spatiotemporal domain. First, we construct the infrared sequence into a holistic spatiotemporal tensor model (STTM) to utilize both spatial and temporal information. Then, based on the sparse-enhanced Tucker decomposition framework, we design a multi-scale energy movement saliency map (MESM) from target motion characteristics. This map serves as a sparse prior, incorporated into the STTM, providing strong decomposition guidance even when the target contrast is weak or overlapped with strong edges. Additionally, we propose a reweighted scheme integrating motion saliency based on the minimum temporal projection of the target component after each iteration, suppressing background clutter and enhancing target separation. Next, for more precise background estimation, we use an accurate low-rank approximation and extend the overlapping group sparse total variation (OGSTV) regularization from 2D to 3D. Compared to traditional variational methods, 3D-OGSTV better distinguishes edges from flat regions, improving background suppression and detection accuracy. Finally, an alternating direction method of multipliers (ADMM) is used for efficient optimization. Experimental results show our approach outperforms state-of-the-art methods, offering better robustness and accuracy in complex scenes.