基于显著性特征的可见光和红外图像融合算法

A Saliency Feature-Based Visible-Infrared Image Fusion Method

摘要: 针对可见光与红外图像在成像机理与信息表达方面存在差异，以及传统融合方法在低照度条件下易出现结构信息不完整、细节表达不足和视觉质量下降等问题，提出一种基于显著性特征的可见光与红外图像融合算法。该方法通过构建互补特征分解模型，将源图像分离为主体结构特征与细节纹理特征，在主体结构特征中使用联合建模像素灰度显著性与多尺度梯度显著性，并引入全变分约束对融合结果进行优化，以提高结构保持能力和整体视觉一致性；在细节纹理特征中使用参数自适应脉冲耦合神经网络(PAPCNN)，根据特征统计特性自动调节网络参数，实现边缘与细节信息的自适应增强。基于LLVIP和M3FD数据集的对比实验结果表明，所提方法在信息熵、特征互信息和视觉信息保真度等评价指标上均优于多种典型融合方法，能够有效提升融合图像的信息表达能力和视觉感知质量。

Abstract: Visible and infrared images exhibit significant differences in imaging mechanisms and information representation. Under low-illumination conditions, traditional fusion methods often suffer from incomplete structural information, insufficient detail preservation, and degraded visual quality. To address these issues, a saliency-based visible and infrared image fusion method is proposed. The proposed method first constructs a complementary feature decomposition model to separate the source images into main structural features and detail texture features. For the structural features, a joint saliency model integrating pixel-intensity saliency and multi-scale gradient saliency is established, and total variation (TV) regularization is introduced to optimize the fusion results, thereby enhancing structural preservation and improving overall visual consistency. For the detail texture features, a parameter-adaptive pulse coupled neural network (PAPCNN) is employed, in which network parameters are automatically adjusted according to the statistical characteristics of the features to achieve adaptive enhancement of edge and fine-detail information. Comparative experiments conducted on the LLVIP and M3FD datasets demonstrate that the proposed method outperforms several representative fusion approaches in terms of information entropy, feature mutual information, and visual information fidelity, effectively improving the information representation capability and perceptual quality of the fused images.