Improved Weighted Nuclear Norm Minimization for Infrared Dim and Small-Target Detection Based on Low-Rank Clustering

In complex infrared scenarios, where prominent edge contours and strong noise interference coexist, effectively suppressing these interfering factors has become the core challenge faced by algorithms for detecting weak and small targets. Current mainstream algorithms generally suffer from a high false alarm rate, mainly due to the interference effect of background clutter. To this end, this paper proposes a Weighted Nuclear Norm Minimization method based on low-rank clustering idea. This method first clusters the image patches adaptively by a similarity measure combining local statistical features and Euclidean distance, which enhances the low-rank property of the background matrix. On this basis, a differentiated weighted nuclear norm penalty is designed for image patches of different categories, which further improves the ability of the model to separate the background from the target. An energy shrinkage mechanism is also introduced to alleviate target energy dilution. Experimental results on six different scenes, compared with six mainstream detection methods and two deep learning methods, using BSF, SNR, and three AUC metrics, demonstrate that the proposed method achieves superior background suppression and detection performance. Although the computational cost is high (80 to 245 s per frame) and currently not suitable for real-time or near-real-time applications, its advantages in background suppression and false alarm control make it particularly suitable for offline or batch processing scenarios, such as post-mission analysis and surveillance data review, where accuracy takes precedence over real-time requirements. This innovative modeling approach provides a new technical solution for weak and small target detection in complex scenarios.