Research on multi-scale decomposition and feature adaptive weight fusion method based on infrared intensity and polarization images

Abstract

In recent years, infrared polarization imaging has gradually become a research hotspot owing to its ability to simultaneously acquire multi-dimensional radiation as well as polarization information and thereby more comprehensively reflect the physical properties of targets. However, it is susceptible to complex environments, causing reduced image contrast and obscured details. Existing multi-scale fusion methods focus mainly on infrared–visible fusion, whereas the fusion of infrared intensity and infrared degree of linear polarization images, which provides complementary thermal-radiation contrast and polarization-sensitive surface detail, has received limited attention. Therefore, a multi-scale decomposition and feature adaptive weight fusion method for infrared intensity and polarization images is proposed. Unlike conventional pyramid-based methods that apply uniform fusion rules across all decomposition levels, the proposed method employs differentiated strategies tailored to distinct frequency components. The source images are first decomposed by Gaussian and Laplacian pyramids into base and detail layers. An adaptive region weight strategy with guided filtering is then applied to the base layers, while an absolute-maximum selection strategy is used for the detail layers. The adaptive weights are computed from local saliency, enabling each image to contribute proportionally to its local informational advantage. The fused layers are then reconstructed into the final fused image. Qualitative and quantitative experiments on two datasets, comparing seven state-of-the-art fusion methods, demonstrate that the proposed method better preserves infrared intensity and polarization information, yielding high contrast and rich detail. This work provides strong support for target detection and recognition in complex environments and demonstrates significant application potential.