Towards CNN-based Level 1 Feature Extraction for Contactless Fingerprint Recognition

Abstract

This work examines the detection of ridge orientation patterns, also referred to as level 1 features, from contactless fingerprint images and their classification. We trained two Convolutional Neural Networks (CNNs) to classify fingerprints based on their ridge orientation patterns. Our models were trained on synthetic data generated by SynCoLFinGer. Afterwards, we conducted various experiments for classifying these patterns and evaluated our trained models on four real-world databases: PolyU CB2CL, ISPFDv1 contactless fingerprint database, and two in-house databases. We report the classification accuracy in terms of Classification Error Rate (CER). We achieved CERs between 28% and 38% considering all samples. Due to the amount of low-quality samples included in the database, we use NFIQ 2.2 to iteratively exclude samples from the databases and report the corresponding CER. We then used NFIQ 2.2 scores to iteratively exclude samples and hence report the impact of low-quality samples. By excluding the lowest scoring 10% of all samples within each database, we achieve CERs of 24% to 35% depending on the databases. While these error rates are still high, they show promise compared to the original values. Although further research is needed to improve results, we show that combining quality-score-based exclusion of images with CNNs trained on synthetic contactless data is a promising method to classify fingerprint patterns.