Abstract:

A light-adaptive image fusion algorithm based on gradient enhancement and text guidance was developed to address the limitations of existing fusion algorithms that cause loss of detailed information， edge degradation， and unclear salient feature under complex lighting environments. Firstly， a feature extraction module based on gradient enhancement and linear spatial equations was constructed to extract global feature with linear computational complexity along with enhancing edge gradient information. Secondly， scene description text was embedded to guide the fusion network to generate fused images of different styles in different lighting environments， so that the robustness of the fusion algorithm in complex lighting environments was improved. Finally， a Gradient Enhanced Fusion Module （GEFM） based on cross-attention mechanism was designed to achieve gradient enhancement and fusion of multimodal information. Experimental results on three benchmark datasets including TNO， MSRS （MultiSpectral Road Scenarios）， LLVIP （Low-Light Visible-Infrared Paired） demonstrate that the proposed algorithm outperforms comparative algorithms such as LRRNet（Low-Rank Representation Network）， CAMF（Class Activation Mapping Fusion）， DATFuse（Dual Attention Transformer Fusion）， UMF-CMGR（Unsupervised Misaligned Fusion via Cross-Modality image Generation and Registration） and GANMcC（GAN with Multi-classification Constraints） in five quantitative metrics. Specifically， the Spatial Frequency （SF） and Visual Information Fidelity （VIF） metrics were improved by 22%， 59%， 61% and 31%， 53%， 37%， respectively. The algorithm effectively reduces edge blurring and ensures that fused images maintain high clarity and contrast under different lighting environments.

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的细节信息丢失、边缘退化和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（GEFM），实现对多模态信息的梯度增强与融合。在3个公开数据集TNO、MSRS（MultiSpectral Road Scenarios）和LLVIP（Low-Light Visible-Infrared Paired）上的实验结果表明，所提算法相较于对比算法LRRNet（Low-Rank Representation Network）、CAMF（Class Activation Mapping Fusion）、DATFuse（Dual Attention Transformer Fusion）、UMF-CMGR（Unsupervised Misaligned Fusion via Cross-Modality image Generation and Registration）和GANMcC（GAN with Multi-classification Constraints）在5种评价指标均有所提高，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31%、53%、37%，有效地减少了边缘模糊，而且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