Towards Regional Fusion for High-Resolution Palmprint Recognition

The existing high resolution palm print matching algorithms essentially follow the minutiae-based fingerprint matching strategy and focus on full-to-full/partial-to-full palm print comparison. These algorithms would face problems when they are applied to forensic palm print recognition where latent marks have much smaller area than full palm prints. Therefore, towards forensic scenarios, we propose a novel matching strategy based on regional fusion for high resolution palm print recognition using regions segmented by major creases features. The matching strategy includes two stages: 1) region-to-region palm print comparison, 2) regional fusion at score level. We first studied regional discriminability of a high resolution palm print under the concept of three regions, i.e., interdigital, hypothenar and thenar, which is the most significant difference between palmprits and fingerprints. Then we implemented regional fusion based on logistic regression at score level using region-to-region comparison scores obtained by a commercial SDK, Mega Matcher 4.0. Significant improvement of recognition accuracy is achieved by regional fusion on a public high resolution palm print database THUPALMLAB. The EER of logistic regression based regional fusion is 0.25%, while the EER of full-to-full palm print comparison is 1%.