[Paper Review] Fingerprint Classification Based on Depth Neural Network

Ruxin Wang, Congying Han+2 more|arXiv (Cornell University)|Sep 18, 2014

Biometric Identification and Security5 references38 citations

TL;DR

This paper proposes a deep learning-based fingerprint classification method using stacked sparse autoencoders to learn orientation field features, followed by softmax regression with a novel fuzzy classification strategy. The approach achieves 98.0% accuracy at a 0.95 probability threshold, significantly outperforming standard methods by reducing misclassification through probabilistic fusion of top two class scores.

ABSTRACT

Fingerprint classification is an effective technique for reducing the candidate numbers of fingerprints in the stage of matching in automatic fingerprint identification system (AFIS). In recent years, deep learning is an emerging technology which has achieved great success in many fields, such as image processing, natural language processing and so on. In this paper, we only choose the orientation field as the input feature and adopt a new method (stacked sparse autoencoders) based on depth neural network for fingerprint classification. For the four-class problem, we achieve a classification of 93.1 percent using the depth network structure which has three hidden layers (with 1.8% rejection) in the NIST-DB4 database. And then we propose a novel method using two classification probabilities for fuzzy classification which can effectively enhance the accuracy of classification. By only adjusting the probability threshold, we get the accuracy of classification is 96.1% (setting threshold is 0.85), 97.2% (setting threshold is 0.90) and 98.0% (setting threshold is 0.95). Using the fuzzy method, we obtain higher accuracy than other methods.

Motivation & Objective

• To improve fingerprint classification accuracy in AFIS by leveraging deep neural networks for unsupervised feature learning.
• To reduce reliance on handcrafted features and singular point detection, which are sensitive to noise and errors.
• To enhance classification robustness by introducing a fuzzy classification method based on multiple probability scores.
• To achieve high-accuracy classification with minimal computational overhead for real-world AFIS deployment.

Proposed method

• Uses the orientation field of fingerprint images as input, bypassing the need for singular point detection.
• Employs a three-layer stacked sparse autoencoder for unsupervised pre-training to learn hierarchical, low-dimensional representations of fingerprint patterns.
• Applies softmax regression on the learned features for multi-class classification into four classes: left loop, right loop, arch, and whorl.
• Introduces a fuzzy classification strategy that combines the top two predicted probabilities to identify and reclassify ambiguous or misclassified samples.
• Sets adaptive probability thresholds (0.85, 0.90, 0.95) to control rejection rate and improve overall accuracy.
• Uses the sum of the first and second highest probabilities as a new condition to recall 30% of previously misclassified samples.

Experimental results

Research questions

• RQ1 Can a deep neural network with stacked sparse autoencoders effectively learn discriminative fingerprint features from orientation fields without relying on singular points?
• RQ2 How does fuzzy classification based on the top two prediction probabilities improve fingerprint classification accuracy compared to standard classification?
• RQ3 What is the optimal probability threshold for balancing accuracy and rejection rate in fingerprint classification?
• RQ4 Can the sum of the two highest class probabilities serve as a reliable criterion to recover misclassified samples?

Key findings

• The proposed method achieves 93.1% classification accuracy using a standard deep network with three hidden layers on the NIST-DB4 dataset.
• By applying fuzzy classification with a threshold of 0.95, the accuracy increases to 98.0%, demonstrating significant improvement over baseline methods.
• The method reduces misclassification by 30% when using the sum of the first and second probabilities as a recovery condition, particularly for ambiguous cases.
• The approach maintains low computational cost while achieving over 99% accuracy when including a secondary classification for suspicious fingerprints.
• The orientation field reconstruction via sparse autoencoders effectively preserves structural patterns, enabling accurate classification even in noisy or ambiguous cases.