Lightweight Convolutional Neural Network Based on Singularity ROI for Fingerprint Classification

Abstract

Fingerprint classification is a significant guarantee for efficient and accurate fingerprint recognition, especially when dealing with one-to-many fingerprint recognition. However, due to large intra-class variability, small inter-class variability, and noise, existing fingerprint classification algorithms still require further improvement in performance and efficiency. In this paper, a Lightweight CNN (Convolutional Neural Network) structure based on singularity ROI (region of interest) is proposed. The experimental results show that the accuracy on testing set of the proposed structure achieves 93%, which is far better than classic non-NN (neural network) classifiers, including RF(Random Forest), KNN (K-Nearest-Neighbor), LR (Logistic Regression), Linear SVM (Support Vector Machine), and RBF (Radial Basis Function) SVM. More momentously, compared with other three CNN structures published in recent years, the proposed structure achieves similar or even better performance with 1/12 to 1/38 parameter scale of other structures, which helps to proceed faster training and testing. Moreover, the proposed CNN model with fewer neurons can achieve better suppression of overfitting and robustness to noise.

Densely Connected Convolutional Network Optimized by Genetic Algorithm for Fingerprint Liveness Detection

“…(i) Firstly, after a series of pre-processing procedures, including normalization, equalization, Gabor enhancement, binarization and thinning [34], the original image is converted into a thinning image. All pixels in thinning fingerprint image are binary (0,255) and all ridgelines are one-pixel thick.…”

“…The dataset includes 10 different sizes of images, ranging from 208 to 1000 in width and from 324 to 1500 in height. Four different operations (flipping up and down, flipping left and right, small-angle rotation, adding Gaussian noise) [34] are applied to augment the training set and enhance the generalization of trained model. However, due to the limited storage capacity of GPU device, only one operation randomly selected from the four augmentation operations is used to produce one augmented image.…”

“…The main contributions of this paper can be summarized as follows: (1) Broad applicability: In this paper, LivDet Datasets (2009, 2011, 2013 and 2015) [3]- [6] are aggregated into one large dataset for model training and testing (Table 2). In order to further improve the generalization of the trained model, four data augmentation operations [34] are utilized to expand the training set. The huge dataset and its augmented dataset ensure the wide applicability of the proposed model.…”

Fingerprint Classification Based on Deep Learning Approaches: Experimental Findings and Comparisons

“…In fact, for efficient one-to-many fingerprint recognition, fingerprint classification is an important step because assessing the target fingerprint only with the same type of fingerprints helps save searching time and the total time of one-to-one recognition. Therefore, fingerprint classification can reduce the execution time of the whole recognition procedure without affecting the performance [15].…”

A Novel Fingerprint Biometric Cryptosystem Based on Convolutional Neural Networks

“…Nguyen et al [17] show the proposed extractor classifies each pixel of a fingerprint image into a category of minutiae with a certain orientation or a non-minutia point, thus obtaining location and orientation information of minutiae simultaneously. In [18], a lightweight CNN structure based on the singularity region of interest is proposed.…”