Abstract: A lighting-adaptive image fusion algorithm incorporating gradient enhancement and text guidance was developed to address the limitations of existing methods that cause loss of detailed information, edge degradation, and unclear salient feature under complex lighting conditions. First, a gradient-enhanced feature extraction module employing linear spatial equations was developed, achieving global feature extraction with linear computational complexity along with enhanced edge gradient information. Second, scene description text embeddings were incorporated to enhance algorithm robustness in complex lighting environments, enabling the fusion network to generate stylistically adaptive outputs under varying illumination conditions. Finally, a cross-attention based fusion module with gradient enhancement was designed to effectively integrate multi-modal information through coordinated gradient optimization. Comprehensive evaluations on three benchmark datasets (TNO, MSRS, LLVIP) demonstrated superior performance over existing methods in five quantitative metrics. Specifically, the spatial frequency and visual information fidelity metrics improved by 22%, 59%, 61% and 30%, 52%, 37%, respectively. Edge blurring was significantly mitigated, with fused images demonstrating enhanced clarity and improved contrast preservation across diverse lighting scenarios.

Key words: Image fusion, Light-adaptive, image gradient enhancement, State Space Model, attention mechanism

摘要： 针对现有融合算法在复杂多变光照环境下存在细节信息丢失、边缘退化、显著信息不明显等问题，提出一种基于梯度增强和文本引导的光照自适应图像融合算法。首先，构建基于梯度增强与线性空间方程的特征提取模块，在实现线性复杂度全局特征提取的同时增强边缘梯度信息；其次，通过嵌入场景描述文本引导融合网络在不同光照环境下生成不同风格的融合图像，提升了融合算法在复杂光照环境下的鲁棒性；最后，构建一种结合交叉注意力机制的梯度增强融合模块，实现对多模态信息的梯度增强与融合。在三个公开数据集TNO、MSRS和LLVIP上的实验结果表明，所提算法相较于对比算法5种评价指标有所提高，其中空间频率（SF）和视觉信息保真度（VIF）指标分别提高了22%、59%、61%和30%、52%、37%，边缘模糊得到了有效地减少，融合图像在不同光照环境下都具有较高的清晰度和对比度。

关键词: 图像融合, 光照自适应, 图像梯度增强, 状态空间模型, 注意力机制