基于双分支注意力机制的指纹纹型分类

PDF下载HTML

Title:

Fingerprint pattern classification based on dual-branch attention mechanism

作者:

赵东越¹, 石磊², 丁锰^1,3

1. 中国人民公安大学 侦查学院, 北京 100038;
2. 中国传媒大学 媒体融合与传播国家重点实验室, 北京 100024;
3. 中国人民公安大学 公共安全行为科学实验室, 北京 100038

Author(s):

ZHAO Dongyue¹, SHI Lei², DING Meng^1,3

1. Department of Criminal Investigation, People’s Public Security University of China, Beijing 100038, China;
2. State Key Laboratory of Media Convergence and Communication, Communication University of China, Beijing 100024, China;
3. Public Security Behavioral Science Lab, People’s Public Security University of China, Beijing 100038, China

关键词:

图像处理; 指纹分类; 双分支结构; 注意力机制; 特征融合; 超参数; 激活函数; 深度学习

Keywords:

image processing; fingerprint classification; dual-branch structure; attention mechanism; feature fusion; hyperparameter; activation function; deep learning

分类号:

TP391.4

DOI:

10.11992/tis.202407005

文献标志码:

2025-1-9

摘要:

针对现有指纹分类算法中存在的低质量指纹难以识别、特征信息提取不充分以及提取过程中信息丢失的问题，提出一种基于双分支注意力机制的指纹纹型分类算法。算法通过提取方向场和进行Gabor滤波的双分支网络进行特征融合，充分利用指纹图像的纹线特征和全局特征；提出的组合激活函数和综合注意力机制模块充分提取卷积分支上的空间特征和通道特征信息，减少提取过程中的信息丢失；设计分支特征融合模块对双分支输出的特征图进行加权，充分融合特征信息；最后引入改进的交叉熵损失缓解样本分布不平衡的问题。实验结果表明，所提算法在自建纹型数据集的4类指纹分类中取得了99.08%的准确率，在准确率、F₁分数和曲线下面积指标方面均优于其他网络模型，验证了本文算法在纹型分类任务上的有效性和优越性。

Abstract:

In order to address the challenges of low-quality fingerprint recognition, insufficient feature extraction and information loss during the extraction process in existing fingerprint classification algorithms, a novel fingerprint pattern classification algorithm based on a dual-branch attention mechanism is proposed. The algorithm employs a dual-branch network for the extraction of orientation fields, which are then subjected to Gabor filtering for the purpose of feature fusion. This approach allows for effective utilisation of both ridge features and global features inherent to fingerprint images. A combination of activation functions and a comprehensive attention mechanism module are proposed to effectively extract spatial and channel feature information from the convolutional branch, thereby reducing information loss during feature extraction. A branch feature fusion module has been devised with the objective of weighting and integrating the feature maps that are output by the dual branches, thereby ensuring comprehensive feature fusion. Finally, an improved cross-entropy loss function is introduced to mitigate the problem of imbalanced sample distribution. The results of the experimental study demonstrate that the proposed algorithm achieves a 99.08% accuracy rate in the classification of four types of fingerprints on a self-built fingerprint pattern dataset, outperforming other network models in terms of accuracy, F₁ score, and area under curve. These findings verify the effectiveness and superiority of the algorithm in fingerprint pattern classification tasks.