Polarization image fusion via analytical attention heads: A multi-scale feature integration framework

Polarization imaging shows great potential for defect detection on highly reflective and low-contrast industrial surfaces. However, existing image fusion algorithms struggle to address the challenges of feature conflicts and polarization noise interference during the polarization fusion process. This paper proposes a polarization image fusion method based on analytical attention heads, aiming to integrate complementary information from different sources while enhancing the prominent features of the main source and suppressing polarization noise. The innovations of this paper are: 1) designing analytical attention heads based on mathematical principles to extract low-level image features such as gradients, textures, information, semantics, and noise; 2) detecting and enhancing prominent features in the main source image to solve the problem of feature loss caused by conflicting feature fusion from different sources; 3) detecting noisy regions in polarization image and reducing their fusion weights to avoid interference from polarization noise. We evaluated our method on both a self-built polarization image dataset and public datasets, and the results demonstrate the advanced nature of our approach. The source code and datasets are publicly available at: https://github.com/FiredTable/DeepFusion.