计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (18): 241-250.DOI: 10.3778/j.issn.1002-8331.2102-0238

• 图形图像处理 • 上一篇    下一篇

基于自注意力深度哈希的海量指纹索引方法

吴元春,赵彤   

  1. 1.中国科学院大学 计算机科学与技术学院,北京 100049
    2.中国科学院大学 数学科学学院,北京 100049
    3.中国科学院大数据挖掘和知识管理重点实验室,北京 100190
  • 出版日期:2022-09-15 发布日期:2022-09-15

Large-Scale Fingerprint Indexing Method Based on Self-Attention Deep Hashing

WU Yuanchun, ZHAO Tong   

  1. 1.School of Computer Science and Technology, University of Chinese Academy of Sciences(UCAS), Beijing 100049, China
    2.School of Mathematical Sciences, University of Chinese Academy of Sciences(UCAS), Beijing 100049, China
    3.Key Laboratory of Big Data Mining and Knowledge Management, Chinese Academy of Sciences, Beijing 100190, China
  • Online:2022-09-15 Published:2022-09-15

摘要: 现有的指纹索引方法大多是基于实数值特征向量,当应用于大规模指纹库时无法避免计算资源与存储空间消耗巨大的问题。为了在海量指纹库中进行高效快速检索并得到实时响应结果,提出了一种全新的基于有监督深度哈希的指纹索引方法。将传统指纹领域知识与自注意力深度哈希模型相结合。传统领域知识用于指纹图像预处理来获取指纹二值骨架图,自注意力深度哈希模型进行特征提取与哈希映射得到二进制编码。其中特征提取模块使用Transformer结构替换卷积神经网络来提取指纹细节特征,此外模型中加入了自动对齐模块并设计了一种STN-AE的结构来辅助训练该模块。最后在NIST4、NIST14、FVC2000、FVC2002、FVC2004等公开指纹数据集上进行了实验,实验结果证实该方法在提高海量指纹库中的检索速度以及降低存储消耗等方面是卓有成效的。

关键词: 指纹索引, 深度哈希, 自动对齐, 指纹骨架图, 自注意力机制

Abstract: Most of the existing fingerprint indexing methods are based on real-valued features, which is compute and storage intensive for large-scale databases. In order to speed up the retrieval process and obtain real-time response results from large-scale databases, a novel fingerprint indexing method based on supervised deep hash is proposed in this paper. This method combines domain knowledge with self-attention deep hash model. The domain knowledge is used for fingerprint image preprocessing to obtain the binary skeleton images, and the self-attention deep hash module maps the binary skeleton images to Hamming space to get binary embeddings for indexing. In the feature extraction module, Transformer structure is used to replace convolutional neural network to extract fingerprint minutiae features. In addition, an automatic alignment module is added to the model and a STN-AE structure is designed to assist in training the module. Experiments conducted on public datasets such as NIST4, NIST14, FVC2000, FVC2002 and FVC2004, have confirmed that this method is effective in improving the retrieval speed and reducing storage consumption.

Key words: fingerprint indexing, deep hashing, automatic alignment, fingerprint skeleton image, self-attention