Identification research based on polarization and Near-Infrared spectrum Dual-Mode image fusion with Multi-Stage overlapping

Abstract

Distinguishing camouflaged fabrics from similar backgrounds has long been a challenge for conventional imaging technologies. Polarimetric hyperspectral imaging offers a novel solution to this problem; however, the massive volume of raw images generated across multiple polarization states and spectral bands increases the complexity of target recognition. This paper proposes a multi-stage image fusion method that integrates existing techniques—Stokes vector analysis, the Normalized Difference Target Index (NDTI), Effective Guided Image Filtering (EGIF), and Retinex-Based Multiphase (RBMP)—into a novel sequential framework, combining dual-mode information from polarization and near-infrared spectra. The integration enhances the differentiation between targets and backgrounds through coordinated multi-stage feature extraction and fusion.Specifically, the method includes polarization image fusion based on the Stokes vector, background suppression using NDTI, detail enhancement through an improved EGIF algorithm, and brightness correction and target recognition via the RBMP algorithm. Validated through indoor bidirectional reflectance distribution function (BRDF) measurements and multi-angle outdoor experiments, the method effectively identifies camouflaged samples against vegetative backgrounds under various incident and imaging angles. Further testing on public datasets confirms its stability and applicability. This approach demonstrates robust performance for target recognition in grassland environments under multi-angle and varying illumination conditions. Further validation in diverse environments is needed before extending to broader scenarios.