Open Access

ARTICLE

DSCAttFuseNet: A Structure–Detail–Luminance Decoupled Network for Low-Light Infrared–Visible Image Fusion

Kezhen Xie, Syed Mohd Zahid Syed Zainal Ariffin^*, Muhammad Izzad Ramli

Faculty of Computer and Mathematical Sciences, Universiti Teknologi MARA, Shah Alam, Malaysia

* Corresponding Author: Syed Mohd Zahid Syed Zainal Ariffin. Email:

Computers, Materials & Continua 2026, 89(1), 104 https://doi.org/10.32604/cmc.2026.083216

Received 31 March 2026; Accepted 13 July 2026; Issue published 13 August 2026

Abstract

Low-light infrared–visible image fusion remains challenging due to severe modality imbalance caused by visible-image degradation under insufficient illumination. In low-light conditions, visible images often suffer from luminance attenuation, blurred details, and amplified noise, whereas infrared images preserve stable structural information but lack texture representation. Existing fusion frameworks commonly perform feature interaction within a shared representation space, which may cause degraded visible responses to be progressively suppressed by dominant infrared structures during fusion. To address this issue, this paper proposes DSCAttFuseNet for low-light infrared–visible image fusion. The proposed framework adopts a structure–detail–luminance decoupled modeling strategy to model infrared structural perception, visible-detail enhancement, and luminance-aware reconstruction through complementary branches, thereby alleviating visible-detail degradation and modality imbalance under low-light conditions. In addition, depthwise separable convolution and efficient channel attention are introduced as compact feature extraction components to reduce redundant feature extraction and strengthen discriminative feature responses. A luminance-guided reconstruction strategy together with multi-objective optimization is further employed to preserve structural consistency, improve visible-detail representation, and maintain balanced luminance representation in fused images. Experiments on the LLVIP dataset and cross-dataset evaluation on the unseen TNO dataset demonstrate that the proposed method achieves competitive performance against representative and recent fusion methods in both qualitative and quantitative evaluations. Ablation studies further verify the effectiveness of the proposed decoupled modeling strategy and compact feature extraction design for low-light infrared–visible image fusion.

---

Keywords

---