Multiframe Infrared Small Target Detection With Frequency-Guided Background Modeling and Saliency-Based Target Enhancement

Infrared small target detection plays a crucial role in both civilian and military applications. Although low-rank and sparse decomposition methods have shown strong capability in background separation, their performance is still limited by inaccurate background modeling and insufficient target enhancement under complex clutter. To address these issues, we propose a novel spatial-temporal detection framework. Consecutive frames are stacked into a tensor, and a Kernelized Tensor Nuclear Norm (KTNN) is introduced to implicitly characterize background low-rank structures in a high-dimensional feature space without explicit rank estimation. A frequency-weighted constraint is further constructed in the DCT domain, where historical background estimates guide current modeling to preserve structural details and improve target–background separability. To enhance target responses, a multi-directional gradient-based saliency prior is developed by integrating directional derivatives with sphering-based statistical features. Meanwhile, an adaptive sparsity reweighting strategy incorporating adjacent-frame feedback is designed to improve target continuity and reduce missed detections. In addition, an inter-component incoherence constraint is imposed to suppress correlations among background, target, and noise. Extensive experiments demonstrate that the proposed method achieves superior detection accuracy, background suppression, and robustness.