Skip to Content
Remote SensingRemote Sensing
  • Article
  • Open Access

18 November 2026

20 Pages

LSUR: Local Statistical and Uncertainty-Guided Refinement for Infrared Small Target Detection

,
,
,
and
1
Beijing Institute of Tracking and Telecommunication Technology, Beijing 100094, China
2
School of Communication Engineering, Xidian University, Xi’an 710071, China
*
Author to whom correspondence should be addressed.
This article belongs to the Section Remote Sensing Image Processing

Abstract

Infrared small target detection (IRSTD) requires the preservation of extremely weak target responses while suppressing structured background clutter. Existing detectors have substantially improved feature representation, yet two issues remain common across architectures: shallow features are usually processed without explicitly measuring local statistical distinctiveness, and high-resolution details are often fused uniformly without considering the reliability of the current prediction. To address these issues, we propose LSUR, a lightweight plug-and-play refinement framework composed of multi-scale local statistical enhancement (MLSE) and confidence–uncertainty-guided refinement (CUGR). MLSE measures the standardized deviation of shallow feature responses within multiple local neighborhoods and adaptively integrates the resulting statistical responses, thereby enhancing locally distinctive target cues with varying background contrast. CUGR derives target-confidence and uncertainty maps from the base prediction and uses them to selectively retrieve high-resolution information for residual logit refinement. The framework requires only one shallow feature, the prediction feature, and the base output, allowing it to be attached to different encoder–decoder IRSTD detectors. Extensive experiments on three public IRSTD datasets and multiple representative detectors demonstrate that LSUR consistently improves detection performance with limited additional computational cost. Specifically, integrating LSUR with SCTransNet improves IoU by 1.68, 0.73, and 1.44 percentage points on NUAA-SIRST, NUDT-SIRST, and IRSTD-1K, respectively. Ablation studies further confirm the effectiveness of local statistical modeling, multi-scale aggregation, and confidence–uncertainty-guided refinement.

Article Metrics

Citations

Article Access Statistics

Article metric data becomes available approximately 24 hours after publication online.