Reliability gated cross-modal attention with shear reparameterized directional reasoning for RGB-thermal salient object detection
Abstract
RGB-Thermal (RGB-T) salient object detection localizes salient regions where appearance cues alone are insufficient, since thermal radiation stays informative under poor illumination. Two limitations persist. Cross-modal attention is applied at a strength fixed in advance, so a momentarily uninformative stream continues to contribute in full, and directional reasoning is confined to the two image axes, because diagonal coverage has required a dense affinity mask of quadratic cost. We present CDRNet, a Confidence and Directional Reasoning Network with dual ResNet-50 encoders. Cross-modal attention is reliability-gated: a coefficient estimated for each modality from spatial entropy and gradient energy scales the transferred context through a gated residual, and the attention scale is derived from the cosine discrepancy between the pooled modality embeddings, so sharpness follows the semantic gap instead of a fixed constant, and balanced inputs recover symmetric fusion. Directional reasoning is shear-reparameterized: a circular shear maps each diagonal onto an axis, so a single axial operator covers four orientations at axial rather than quadratic cost, a per-pixel gate exposes the learned anisotropy, and a weight-shared second stage refines the result without additional parameters. A boundary gated decoder returns edge evidence to the representation through multiplicative modulation under auxiliary supervision, and its cross scale aggregator mixes two dilation rates per branch under context-dependent weights. On VT821, VT1000, and VT5000, CDRNet attains S-measure values of 0.908, 0.949, and 0.918, surpasses 19 recent methods on all metrics, and ablations isolate each contribution.
Related
- YOLO-RMF: A Real-Time Surface Defect Detection Model for Hot-Rolled Strip Steel with Reparameterized Convolution and Axial Attention
- EDGRU: edge-guided dual graph reasoning U-Net with spatial graph attention for infrared small target detection
- Cloud–Edge Collaborative Hardware Sensing with Dynamic Reliability Estimation and Evidence-Constrained Large Language Model Reasoning
- Retrieval-augmented reasoning reshapes collective reliability in text-based radiology question answering
- Dynamic heterogeneous graph attention network for spatiotemporal correlation modeling of mine microseismic events and intelligent rockburst risk level reasoning
BibTeX
@article{othman2026reliability,
title = {{Reliability gated cross-modal attention with shear reparameterized directional reasoning for RGB-thermal salient object detection}},
author = {Kamal M. Othman and Abdulfattah Noorwali and Manal Abdullah Alohali and Inam Ullah and Mi Young Lee},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-72418-w},
url = {https://www.nature.com/articles/s41598-026-72418-w},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Reliability gated cross-modal attention with shear reparameterized directional reasoning for RGB-thermal salient object detection
AU - Kamal M. Othman
AU - Abdulfattah Noorwali
AU - Manal Abdullah Alohali
AU - Inam Ullah
AU - Mi Young Lee
PY - 2026
DA - 2026/10/01
JO - Scientific Reports
DO - 10.1038/s41598-026-72418-w
UR - https://www.nature.com/articles/s41598-026-72418-w
AB - RGB-Thermal (RGB-T) salient object detection localizes salient regions where appearance cues alone are insufficient, since thermal radiation stays informative under poor illumination. Two limitations persist. Cross-modal attention is applied at a strength fixed in advance, so a momentarily uninformative stream continues to contribute in full, and directional reasoning is confined to the two image axes, because diagonal coverage has required a dense affinity mask of quadratic cost. We present CDRNet, a Confidence and Directional Reasoning Network with dual ResNet-50 encoders. Cross-modal attention is reliability-gated: a coefficient estimated for each modality from spatial entropy and gradient energy scales the transferred context through a gated residual, and the attention scale is derived from the cosine discrepancy between the pooled modality embeddings, so sharpness follows the semantic gap instead of a fixed constant, and balanced inputs recover symmetric fusion. Directional reasoning is shear-reparameterized: a circular shear maps each diagonal onto an axis, so a single axial operator covers four orientations at axial rather than quadratic cost, a per-pixel gate exposes the learned anisotropy, and a weight-shared second stage refines the result without additional parameters. A boundary gated decoder returns edge evidence to the representation through multiplicative modulation under auxiliary supervision, and its cross scale aggregator mixes two dilation rates per branch under context-dependent weights. On VT821, VT1000, and VT5000, CDRNet attains S-measure values of 0.908, 0.949, and 0.918, surpasses 19 recent methods on all metrics, and ablations isolate each contribution.
SN - 2045-2322
ER -