POLAR: Polarization-Aware Object Localization and Recognition Framework for Concealed Threats in Passive Millimeter-Wave Images

As a subband of the terahertz spectrum, passive millimeter-wave (PMMW) imaging is ideal for human security screening due to nonionizing radiation and clothing-penetrating capabilities, but its poor image quality limits concealed threat recognition accuracy. To address this, we exploit polarization—a fundamental property of electromagnetic waves—as a rich source of complementary information to differentiate between the human body and concealed objects. This article proposes a polarization-aware object localization and recognition (POLAR) framework to construct the deep learning framework for PMMW imaging detection. The core innovation is a masked deformable convolution mechanism that integrates physically derived polarization masks to guide the spatial attention. These masks are computed from Stokes parameters extracted from four linear polarization (0°, 45°, 90°, 135°) brightness temperature images, effectively highlighting key regions with high contrast in each polarization modality. The POLAR adopts a multiscale processing pipeline, where multipolarization features at each scale are fused using dual-dimensional weights from both channel and spatial domains, and the recognition results are obtained based on the multiscale fused features. Compared with existing state-of-the-art methods, our method achieves the best overall performance while simultaneously maintaining high accuracy in comprehensive identification and high computational efficiency. The work demonstrates that embedding polarization physics into a deformable convolution framework significantly enhances the accuracy and interpretability of concealed threat detection in low-quality PMMW imagery.