Infrared Small Target Detection via tensor $L_{2,1}$ norm minimization and ASSTV regularization: A Novel Tensor Recovery Approach
Jiqian Zhao, An-Bao Xu
TL;DR
The paper tackles infrared small target detection in sequences by addressing the computational burden of nuclear-norm based methods. It introduces ASSTV-TLNMTQR, a fast tensor recovery framework that combines tensor $L_{2,1}$-norm minimization via tensor QR decomposition with ASSTV regularization, enabling simultaneous background suppression and target extraction through a sliding-window tensor decomposition into background, target, and noise. The approach uses ADMM to solve a joint optimization that enforces a low-rank background via $L_{2,1}$-norm, sparse targets, and asymmetric spatial-temporal regularization, achieving significant speedups over prior methods while maintaining accuracy. Experiments on six ISTD sequences with 3D ROC metrics demonstrate improved detection performance and about 25% faster processing compared with prior ASSTV-based methods, highlighting practical impact for real-time infrared detection scenarios.
Abstract
In recent years, there has been a noteworthy focus on infrared small target detection, given its vital importance in processing signals from infrared remote sensing. The considerable computational cost incurred by prior methods, relying excessively on nuclear norm for noise separation, necessitates the exploration of efficient alternatives. The aim of this research is to identify a swift and resilient tensor recovery method for the efficient extraction of infrared small targets from image sequences. Theoretical validation indicates that smaller singular values predominantly contribute to constructing noise information. In the exclusion process, tensor QR decomposition is employed to reasonably reduce the size of the target tensor. Subsequently, we address a tensor $L_{2,1}$ Norm Minimization via T-QR (TLNMTQR) based method to effectively isolate the noise, markedly improving computational speed without compromising accuracy. Concurrently, by integrating the asymmetric spatial-temporal total variation regularization method (ASSTV), our objective is to augment the flexibility and efficacy of our algorithm in handling time series data. Ultimately, our method underwent rigorous testing with real-world data, affirmatively showcasing the superiority of our algorithm in terms of speed, precision, and robustness.