Abstract
We propose a new approach to matching mechanism for palmprint recognition byintroducing a Relative Similarity Metric (RSM) that enhances the robustness anddiscriminability of existing matching frameworks. While conventional systemsrely on direct pairwise similarity measures, such as cosine or Euclideandistances, these metrics fail to capture how a pairwise similarity compareswithin the context of the entire dataset. Our method addresses this byevaluating the relative consistency of similarity scores across up to allidentities, allowing for better suppression of false positives and negatives.Applied atop the CCNet architecture, our method achieves a new state-of-the-art0.000036% Equal Error Rate (EER) on the Tongji dataset, outperforming previousmethods and demonstrating the efficacy of incorporating relational structureinto the palmprint matching process.