Extraction of the Features of Fingerprints Using Conventional Methods and Convolutional Neural Networks
Karthik E. M. V. Naga, Gopal Madan · 2021
Fingerprints are one of the most common biometric identifiers. Until recently, the conventional image processing methods were being used in extracting the features of the fingerprints for the fingerprint classification problem. However, with the rise of artificial intelligence, deep learning models such as the Convolutional Neural Networks (CNNs) have shown promising results in image classification problems. In this paper, we explain why CNNs are performing better by visualizing the features learned by its convolutional layers, and comparing them with the fingerprints' features extracted by estimating the local orientation map and detecting the singular regions. A 17-layer CNN model is proposed, which obtains a classification accuracy of 92.45% on the NIST-DB4 fingerprint dataset. We conclude that the first two convolutional layers are learning features that are similar to the ones obtained after using the above techniques, while the remaining layers are learning abstract and more complex features that are class-specific. This explains why the deep learning models are performing better. The results are promising as they bring us one step closer in demystifying the inner functioning of the CNNs.