The Application of Convolutional Neural Networks for Fingerprint Recognition: A Comparative Analysis

The use of convolutional neural networks (CNNs) in the domain of biometric identification is examined in this paper. On a well-known fingerprint evaluation dataset, the three most widely used networks AlexNet, GoogLeNet, and ResNet were tested in the positive identification scenario. In order to improve interclass discrimination and coherence of input data, image enhancement and region of interest segmentation were used to remove inconsistent regions that are typically present on the image peripheral, generating spurious features that negatively affect the neural network learning rate. The fingerprint database prepared in this technique represents the input data of the identification system entirely based on the capabilities of the CNN, allowing direct comparison of networks performances and selection for further implementation. As a result, trained CNNs can be used as a feature extraction module in biometric cryptosystems or robust authentication systems. However, testing results reveal that the proposed technique outperforms many biometric authentication systems, with 97% accuracy rate.