| 摘要: | 近年來，像筆記型電腦、智慧型手機及平板電腦的各種行動裝置已經成為人們生活中的一部份，而保證在這些行動裝置上資料的安全是必要的。指紋是其中一種足以識別個別身分的生物特徵，被用於各種場合，包括付款、銀行、出勤及個人財物。然而，因為行動裝置上的指紋感測器大小只有10 x 10 mm2，所得的影像為局部指紋，像紋路、特徵點及毛孔這類的資訊可能不足，在行動裝置上辨識指紋影像是困難的。
本篇論文提出一種新的局部指紋辨識方法。首先，用來識別身分的指紋特徵由AKAZE演算法計算得出，這種局部特徵不受指紋影像的雜訊、尺度及方向影響。其次，指紋影像的比對採用滑動視窗的方式，這種方式能徹底地進行完整指紋影像對局部指紋影像的比對。最後，比對分數由人工神經網路計算得出，比起啟發式方法，人工神經網路可以在缺少規則的情形下辨別多樣化的資料。
本篇論文在FVC2002資料庫上進行實驗並評估演算法效果，結果顯示所提出的方法滿足生物識別技術的相關需求。所得的EER及FRR@FAR 1/50000皆小於9%，最好的結果出自DB1，EER達4.95%，FRR@FAR 1/50000達6.06%。然而，演算法所設定的指紋影像解析度為184 x 184 pixels，尚無法辨識較低解析度的指紋影像。未來的研究會以適應各種不同的影像解析度為目標進行改進，而在電腦視覺領域中，深度學習方法被用來解決許多複雜的問題，應該能使演算法在不同的影像解析度下保持水準。
;In recent years, mobile devices such as laptops, smartphones, and tablets have become an integral part of human activities. Accordingly, we need a method to ensure data confidentiality on mobile devices. The fingerprint is one of the unique biometric traits used to authenticate individuals. It is widely used for security issues, including payments, banking, attendance, and securing belongings. However, recognizing the fingerprint image on a mobile device is difficult since the fingerprint reader is only 10 x 10 mm2 in size. The reader captures only some parts of a whole fingerprint. This result contains only a small amount of information, such as ridges, minutiae, and pores.
We proposed a novel methodology for partial fingerprint recognition. The fingerprint feature was represented using the local feature-based AKAZE. These features were selected because they could maintain the fingerprint image′s noise, scale, and orientation. Moreover, we formulated the standard of matching tasks using a sliding window. This sliding window made it possible to compare full images and partial queries comprehensively. As part of this approach, we also replaced the heuristic method of calculating the matching rate with neural networks. A neural network had proven to be able to distinguish between a variety of data without the need for a lot of rules.
As validation, we experimented with our method using an instance of the FVC2002 database. Our method achieves adequate results in terms of biometric evaluation. These values of EER and FRR@FAR 1/5000 are both less than 9%. The highest success rate was recorded in DB1, EER reached 4.95%, and FRR@FAR 1/50000 reached 6.06%. However, these methods only optimize for images of 184x184 pixels. We cannot yet recognize fingerprints when the resolution is reduced. As future research, we must ensure our method work in various image resolutions. Deep learning can help sustain performance across resolutions. As far as computer vision is concerned, deep learning has been proven to be successful in solving complex problems. |
