基于深度卷积神经网络的指纹纹型分类算法

传统指纹纹型分类算法的准确率直接受到相应特征提取算法的影响.在海量指纹库中，同类纹型指纹形态变化明显增大，不同类纹型界限变得模糊，仅通过人工定义的特征进行分类很难适应全部指纹数据.为解除纹型分类问题与人工定义的特征提取问题的耦合，提出一种直接在指纹原图上进行纹型识别的算法.利用卷积神经网络自动特征提取的能力从大量指纹数据中学习得到纹型特征，并通过对训练数据的设计使网络能够适应指纹的多样性，提升算法的鲁棒性.此外，多尺度网络模型平均方法使分类准确性得到进一步提升.在国际公开指纹数据集NIST DB4上测得纹型四分类准确率达94.2%.

Fingerprint pattern classification algorithm based on deep convolutional neural networks

The accuracies of traditional fingerprint pattern classification algorithms rely heavily on the corresponding feature extraction algorithms. Further more, the within-class variance of fingerprint patterns increases while the between-class variance decreases in large-scale database. So it is difficult for hand-designed features to suit with all fingerprint data. In order to remove the coupling with hand-designed feature extraction algorithms, we propose an approach to directly recognize patterns in raw fingerprint images. It takes advantage of the automatic feature extraction ability of convolutional neural networks to learn patterns from large amount of images. The training data are carefully designed to fit the variety of fingerprints and to improve the robutness. Meanwhile, the accuracy is further improved by averaging multi-scale models. In our experiment, an accuracy of 94.2% for four-class classification has been achieved in the international opening fingerprint dataset NIST DB 4. Our algorithm surpasses many classical algorithms, and it is both practical and meaningful.