Deep Unfolding with Target Confidence and Cross-Stage Memory for Infrared Small Target Detection

Infrared small targets occupy few pixels and often exhibit weak contrast, making them difficult to distinguish from bright edges, background clutter, and noise. Deep unfolding provides an interpretable framework for background–target decomposition, yet weak target responses may be absorbed into background estimates or suppressed during sparse updates. We propose a Target Confidence and Memory Network (TCMNet) that incorporates target confidence and cross-stage error memory into a multistage deep unfolding framework. A confidence-guided multiscale background modeling module combines smoothing priors at different scales with learned background representations. In later stages, target confidence derived from intermediate target estimates reduces the influence of potential targets on background estimation. A transform-domain sparse update module employs learnable transforms and channel-wise learnable thresholds to suppress clutter, together with local target protection to limit the attenuation of weak responses. To exploit information from preceding stages, a cross-stage error memory module integrates current target responses, inter-stage response changes, and reconstruction residuals into a recurrent state. This state produces bounded corrections to the target estimate before it is passed to the next stage. Target estimates are constrained to be nonnegative, while an auxiliary reconstruction loss penalizes discrepancies between reconstructed and observed images. A final prediction head combines target estimates, multiscale local contrast cues, and the reconstruction residual to produce a target probability map. Experiments on NUDT-SIRST and a local 1:1 partition of MDvsFA demonstrate the effectiveness of TCMNet, achieving mIoU scores of 91.71% and 90.31%, F1 scores of 95.68% and 94.91%, and detection probabilities of 98.41% and 96.84%, respectively. These results highlight the ability of the proposed framework to preserve weak target responses while suppressing background interference. By integrating explicit background–target decomposition with adaptive sparse updating and cross-stage error correction, TCMNet provides a structured approach to infrared small target detection in complex scenes.

Authors

Institutions

Publication Details

Journal
Remote Sensing
Published
2026-11-19
DOI
https://doi.org/10.3390/rs18223973
Primary Topic
Infrared Target Detection Methodologies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deep Unfolding with Target Confidence and Cross-Stage Memory for Infrared Small Target Detection

Xiaoguang Shao, Yiming Zhao, Xiang Li, Yiwen Wang et al.
Remote Sensing
Infrared Target Detection Methodologies
article

Deep Unfolding with Target Confidence and Cross-Stage Memory for Infrared Small Target Detection

Xiaoguang Shao, Yiming Zhao, Xiang Li, Yiwen Wang, Wei Zheng, Jingyu Li
article en

Abstract

Infrared small targets occupy few pixels and often exhibit weak contrast, making them difficult to distinguish from bright edges, background clutter, and noise. Deep unfolding provides an interpretable framework for background–target decomposition, yet weak target responses may be absorbed into background estimates or suppressed during sparse updates. We propose a Target Confidence and Memory Network (TCMNet) that incorporates target confidence and cross-stage error memory into a multistage deep unfolding framework. A confidence-guided multiscale background modeling module combines smoothing priors at different scales with learned background representations. In later stages, target confidence derived from intermediate target estimates reduces the influence of potential targets on background estimation. A transform-domain sparse update module employs learnable transforms and channel-wise learnable thresholds to suppress clutter, together with local target protection to limit the attenuation of weak responses. To exploit information from preceding stages, a cross-stage error memory module integrates current target responses, inter-stage response changes, and reconstruction residuals into a recurrent state. This state produces bounded corrections to the target estimate before it is passed to the next stage. Target estimates are constrained to be nonnegative, while an auxiliary reconstruction loss penalizes discrepancies between reconstructed and observed images. A final prediction head combines target estimates, multiscale local contrast cues, and the reconstruction residual to produce a target probability map. Experiments on NUDT-SIRST and a local 1:1 partition of MDvsFA demonstrate the effectiveness of TCMNet, achieving mIoU scores of 91.71% and 90.31%, F1 scores of 95.68% and 94.91%, and detection probabilities of 98.41% and 96.84%, respectively. These results highlight the ability of the proposed framework to preserve weak target responses while suppressing background interference. By integrating explicit background–target decomposition with adaptive sparse updating and cross-stage error correction, TCMNet provides a structured approach to infrared small target detection in complex scenes.

Remote SensingVol. 18(22)
Henan University of Technology (CN), National Space Science Center (CN)
Infrared Target Detection Methodologies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon