Feature-Prior-Guided Improved Infrared/Visible Image Fusion Algorithm

Multimodal image fusion offers significant value in applications like military detection and medical imaging. However, traditional approaches often lack prior knowledge guidance or clear fusion principles, limiting interpretability. To address this, we propose an improved infrared/visible image fusion algorithm guided by feature priors. Our method comprises three components: a feature prior extraction module, an attention fusion network, and a loss function. The feature prior extraction module ranks the importance of image features—including color, texture, shape, local characteristics, frequency domain, and depth—using a combined machine learning algorithm and SHAP value analysis, providing an importance prior. Building on the SeAfusion network, the attention fusion network then integrates channel and spatial attention mechanisms, uniquely guided by these traditional feature priors for deep feature fusion. The loss function combines content loss (encompassing intensity, gradient, and perceptual losses) and semantic loss to jointly optimize the network. Evaluated on the MSRS dataset, the proposed algorithm outperforms SeAfusion in metrics including PSNR, CC, and SSIM. Furthermore, its feature visualizations demonstrate enhanced interpretability, effectively validating the method’s superiority.