Optimizing Fingerprint Identification: CNNs With Raw Images Versus Handcrafted Features for Real-Time Systems

Fingerprint identification is a cornerstone in various domains such as security, forensics, and authentication. Despite progress, existing systems still face challenges with noise, database differences, and real-time speed. This study investigates the balance between accuracy and computational efficiency(thereby speed) by comparing two approaches: training a Convolutional Neural Network (CNN) with raw fingerprint images and training a CNN using handcrafted fingerprint features. Comprehensive evaluations were conducted across multiple benchmark datasets FVC2000, FVC2002, FVC2004, and NIST SD27, using both single-run and multi-run experiments. Single-run experiments assessed performance across epochs; multi-run cross-validation on FVC2000 ensured consistent and reliable results. A similarity-based matching mechanism with an optimal threshold of 0.8, determined through Receiver Operating Characteristic (ROC) curve analysis, was employed for classification. Experimental results on NVIDIA DGX A100 GPU demonstrate that CNN training with handcrafted features consistently achieves superior accuracy (99.80%) compared to CNN training with raw images approach accuracy (95.79%) on the FVC2000 dataset, while significantly reducing computation time from 1465 to 203 seconds. This performance was consistent across all experimental runs. Similar trends appeared across datasets, with handcrafted features reaching 96-99% accuracy versus 91-96% for raw images data. Statistical analyses (t-tests, ANOVA, and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</i>-statistics) confirmed the significance of these performance differences (p < 0.01) and demonstrated the model’s reliability, robustness, and consistency. This proposed method outperforms recent methods in accuracy and speed, validated by ROC, precision-recall, F1-score, and accuracy metrics in real-world scenarios. Our approach provides experimental confirmation that CNNs fed with features provide improvements in accuracy and speed of classification.