Biometric Fingerprint Verification with Siamese Neural Network and Transfer Learning

In this research, a fingerprint verification model, equipped with a Siamese Neural Network structure with MobileNetV2 and ConvNeXt models, was established in developing a robust fingerprint verification system using transfer learning. The algorithm is learned using a custom-made fingerprint database with image pairs dubbed as similar or different that facilitates the model to recognize the discriminative characteristics of biometric comparison. The Siamese structure enables the system to have generalization of unseen fingerprint classes without retraining again to be scalable and adaptable during applications in the real world. This MobileNetV2 performed very well with an accuracy of 96.35 and F1-scores of 0.96 on both classes and was very capable of differentiating between similar and dissimilar fingerprints. However, it was surpassed in accuracy (98.61% higher) and balanced F1-scores (0.99 higher) and on the generalization and classification error superiority by ConvNeXt. The overall scores of the two models were 1.00 with an area under the receiver operating characteristic curve, indicating that they were in a perfect state of class separability. ConvNeXt also demonstrated a cleaner convergence curve during the training process, a fact that further confirmed its superiority in cases of dynamic biometric verification. In general, the findings indicate the viability of deep learning based Siamese networks and elaborate convolutional networks in terms of scalable and precise fingerprint crack verification solutions.