A Comparative Analysis of Fingerprint Classification: Traditional Machine Learning with KMCG vs. Deep Learning Approaches

This study presents a novel comparative analysis between traditional machine learning (ML) methods—enhanced through the Kekre’s Median Codebook Generation (KMCG) technique—and modern deep learning approaches for fingerprint classification. For the traditional ML pipeline, models including Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), and XGBoost were evaluated using features extracted via the KMCG algorithm. Notably, this study explores the under-researched application of KMCG in fingerprint classification, revealing that a configuration with a codebook size of 2 and window size of (16, 16) yields the best performance at $77.20 \%$ feature-level accuracy. Among these models, SVM attained the highest classification accuracy at $70.2 \%$. On the other hand, the models that applied the transfer learning concept, MobileNetV2, InceptionV3, and DenseNet201, showed much higher efficiency. In this paper, MobileNetV2 achieved a peak classification accuracy of $91.72 \%$ and peak precision of $95.92 \%$, indicating that deep feature hierarchies and pre-trained representations are beneficial in identifying the complex and abstract geographical patterns of fingerprints. The studies presented indicate that KMCG can be helpful in further developing biometric feature extraction, as well as highlighting the importance of deep learning for achieving high classification accuracy of fingerprints. This work helps to understand the trade-off between reporting results in traditional ML and deep learning, and is useful to develop modern biometric systems.