基于HRF-Net的指纹细节点及汗孔统一提取方法

A Unified Method for Minutiae and Sweat Pore Extraction Using HRF-Net

基于多级特征(例如细节点、汗孔等)融合的指纹识别技术,大大提高了指纹识别系统的安全性和鲁棒性.然而,目前基于高精度指纹的识别技术,几乎都是基于第三级特征中的汗孔特征,而忽略了指纹图像中的其他重要特征.针对这一问题,本文首次提出一种指纹特征提取方法,能够实现在高精度指纹图像上同时提取不同层级特征,包括二级的细节点特征和三级的汗孔特征.本文设计了High-Resolution Fingerprint Net(HRF-Net)作为特征提取模型,利用指纹图像生成细节点与汗孔的热力图,再通过滑动窗口算法处理得到特征点坐标.在HRF-Net模型中,通过引入中继输出结构以实现汗孔和细节点特征的解耦,利用由粗到细的阶段式监督以兼顾网络对不同层级特征的学习,在网络中加入shuffle unit模块减少模型计算复杂度,实现了对指纹不同层级特征高效准确的提取.实验结果表明,本文提出的特征统一提取方法在汗孔的提取上真阳率(RT)达到了96.59%,比目前取得最好性能的Judge CNN提高了3.45%;在细节点的提取上真阳率(RT)达到了81.93%.同时,我们在对汗孔和细节点单独提取上也达到了最好的结果,以衡量提取综合性能的F1-score作为评价指标,模型提取汗孔的F1-score达到了95.83%,比Judge CNN提高1.48%.我们利用所提取的特征在指纹匹配数据集上的指纹图像进行匹配实验,在等错误率(Equal Error Rate,EER)上达到了5.39%,相比传统方法下降7.02%.结果表明,本文的方法在汗孔和细节点的提取性能以及匹配结果上都达到了目前最佳水平.

Fingerprint features have three levels of different characteristics,namely first-level features(shape and direction of ridges,etc.),second-level features(minutiae,etc.),and thirdlevel features(pores,etc.).Traditional fingerprint recognition systems typically rely only on first and second-level features,particularly minutiae.Fingerprint recognition technology based on the fusion of multi-level features(such as minutiae,pores,etc.)has greatly improved the security and robustness of fingerprint recognition systems.Sweat pores are a crucial aspect of high-resolution fingerprint image recognition.However,current high-resolution fingerprint recognition technology predominantly focuses on the sweat pore feature as a third-level characteristic,often overlooking other significant features present in fingerprint images.To address this issue,this article introduces the High-Resolution Fingerprint Net(HRF-Net)as a feature extraction model,which utilizes fingerprint images to generate heat maps of minutiae and sweat pores.These heat maps are then processed using a sliding window algorithm to obtain the coordinates of feature points.In the HRF-Net model,the introduction of intermediate outputs structure allows for the separation of sweat pore and minutiae features.Additionally,a staged supervision approach,starting from coarse to fine,is employed to ensure the network learns different levels of features effectively.To reduce computational complexity,a shuffle unit module is incorporated into the network,enabling efficient and accurate extraction of fingerprint features at various levels.By generating heat maps of minutiae and sweat pores,it captures the intricate details of the fingerprint,enabling a more comprehensive representation of the features.The introduction of intermediate outputs structure allows for the disentanglement of sweat pore and minutiae features,contributing to a more focused and refined feature extraction.Additionally,the staged supervision approach ensures that the network learns the different levels of features progressively,enabling a holistic understanding of the fingerprint image.Furthermore,the incorporation of the shuffle unit module reduces the computational complexity of the model.The combination of these techniques results in a highly efficient and accurate fingerprint feature extraction model.Experimental results show that our proposed unified extraction method achieves a true positive rate of 96.59%in pore extraction,which is 3.45%higher than the best-performing Judge CNN.The true positive rate of minutiae extraction reaches 81.93%.At the same time,we also achieved the best results in separate extraction of pores and minutiae.The F1-score for the extraction of pores reaches 95.83%,which is 1.48%higher than that of Judge CNN.We use the extracted features to conduct matching experiments on the fingerprint matching dataset,and achieve an equal error rate(EER)of 5.39%which is 7.02%reduction compared to traditional methods.These results indicate that our proposed HRF-Net model delivers superior performance in pore and minutiae extraction,as well as matching accuracy.By leveraging the extracted features,our method significantly enhances the efficiency and reliability of fingerprint recognition systems.The HRF-Net model holds great potential for applications in biometric security and forensics,offering a promising solution for high-resolution fingerprint feature extraction and matching.