A registration method for infrared and visible images based on normalized shape description

Traditional image feature matching techniques mainly rely on intensity- or gradient-based information for feature detection and description. However, these approaches are highly susceptible to nonlinear radiometric distortions (NRD) prevalent between infrared and visible images, often resulting in poor registration performance. To overcome these limitations, we propose a novel infrared and visible image registration framework named Normalized shape description (NSD) based registration. The NSD method emphasizes the extraction of stable shape features rather than conventional point features, thus enhancing the robustness of the registration. Specifically, the approach employs image segmentation to identify shape features within the images. It then constructs scale- and rotation-invariant descriptors utilizing normalized radial distances, effectively eliminating the dependence on intensity and gradient information. Subsequently, cosine similarity is employed to perform a global comparison of edge point sets derived from shape features, facilitating the reliable establishment of corresponding point pairs. Experimental validation was conducted on 110 image pairs selected from six infrared and visible datasets, with comparisons against eight state-of-the-art methods. The results demonstrate that NSD significantly improves registration performance under conditions of nonlinear radiometric distortion, achieving superior metrics in registration success rate, correct matching rate, and the number of accurate correspondences.