Polarization image fusion based on structure and detail decoupling guided by Edge Prior

Polarization image fusion aims to enhance texture of the intensity image ( \(S_0\) ) by integrating the material-sensitive cues contained in the Degree of Linear Polarization (DoLP) image. However, existing fusion methods struggle to achieve a controllable balance between texture enhancement and structural preservation, which often leads to either \(S_0\) -copying degeneration or the injection of invalid information. In this paper, we propose EPFuse, an asymmetric structure–detail decoupling framework for polarization image fusion. First, the proposed framework decomposes each modality into base-structure and detail streams, associated with an explicit decoupling regularizer to reduce representation leakage between the two streams. Then, to selectively inject polarization cues, we design a parameter-free, bounded DoLP Edge Prior module that acts as an additive spatial bias in the learned competition between the \(S_0\) and DoLP detail streams, while \(S_0\) remains the dominant carrier of global structure and color. A composite objective jointly enforces structural fidelity to \(S_0\) , smoothness, and region-aware high-frequency consistency. By doing so, EPFuse promotes polarization-detail enhancement in gradient-salient regions while reducing the propagation of moderate random perturbations. In this work, we also construct a real-world dataset consisting of aligned \(S_0\) -DoLP pairs, which facilitates systematic evaluation of polarization image fusion in diverse outdoor scenes. Experiments on public datasets and our dataset demonstrate that EPFuse improves both perceptual quality and downstream detection accuracy.