【摘要】 传统红外偏振图像融合方法严重依赖于偏振度对噪声高度敏感的特征，在处理低信噪比偏振数据时，往往导致特征提取失败和融合效果不佳，为此，提出了一种两阶段融合方法。首先，为摆脱对传统偏振特征的依赖，基于斯托克斯理论和正交偏振矢量模型构建了一种名为“结构化去噪正交偏振矢量”（S-OPV）的新型特征描述，通过数学变换与结构化去噪，能够从原始偏振数据中直接提取出稳定且轮廓清晰的目标特征；其次，为实现两种模态信息的高效融合，设计了一个基于边缘感知的双分支卷积神经融合网络，该网络能够自适应地保留红外强度的纹理细节，同时利用注意力引导机制将S-OPV提供的强偏振特征精确、合理地注入到融合结果中。这种红外偏振图像融合方法在噪声抑制、轮廓保持和目标显著性提升方面均表现出优越的性能。更多还原

【Abstract】 Conventional Infrared polarization image fusion methods heavily rely on the characteristics which Degree of Polarization(DoP)are highly sensitive to noise,often leading to failure in feature extraction and poor fusion results when processing low-SNR polarization data. To overcome this,a two-stage fusion method is proposed. First,to eliminate dependence on traditional polarization features,a novel feature descriptor called “Structured-Denoised Orthogonal Polarization Vector(S-OPV)” is constructed based on Stokes Theory and orthogonal polarization vector model. Through mathematical transformation and structured denoising,this method can directly extract stable and well-defined target features from raw polarization data. Second,to achieve an efficient fusion of two modes information,an edge-aware dual-branch convolutional neural fusion network is designed. This network can adaptively retains the texture details of infrared intensity while precisely and reasonably injecting the strong polarization features provided by S-OPV into the fused image using an attention-guided mechanism. The proposed infrared polarization image fusion method demonstrates superior performance in noise suppression,contour preservation,and target saliency enhancement.更多还原