基於壓縮領域之指紋辨識系統

Abstract

隨著3C產品與個人資料的大量成長，掌管使用權的身分辨識技術逐漸受到重視，其中利用生物特徵進行辨識的作法，已逐漸取代傳統知識型驗證成為主流趨勢，例如在智慧型手機的解鎖系統中，指紋辨識已佔有廣大的市場，而新興的屏下指紋辨識系統，能夠直接將指紋感測器安裝於螢幕下方，因此不需要額外的空間放置指紋感測器，不但使用起來更方便，也提供全面屏手機指紋辨識的做法。然而，屏下指紋易受到顯示面板與相關電路的雜訊影響，導致圖像品質不佳，要使用傳統局部特徵點的辨識作法，就須要先透過高運算與費時的指紋重建方法，以提升指紋圖像品質，此過程對於講求低運算複雜度的終端裝置是較不友善的。 針對上述問題，本文利用全域方向場(Orientation Field)提取特徵，觀察角度分布以建立頻率向量(Frequency vector)，由於方向場對於雜訊有較高的容忍度，因此高複雜度的指紋重建方法將可以被省略。為了更進一步提升辨識能力，我們使用三元組損失(Triplet loss)來訓練類神經網路模型，將原先頻率向量變換為一個轉換向量(Transformed vector)，使之拉近自身特徵間距離並推離其他資料的特徵。最後利用比對基礎的辨識方法(Matching-based method)進行指紋辨識，其中表決決策法(Vote decision method)能使自身與其他指紋間的辨識分數有更明顯的差距，以此獲得更好的辨識結果。綜合以上所述，本文設計了一個基於全域向量場的指紋辨識作法，並利用類神經網路增強萃取特徵，最後透過比對基礎方式進行指紋辨識，此作法不須要高複雜度的指紋重建過程，將更適合終端裝置使用。

With the massive growth of 3C products and personal data, user identification that controls the access right has gained a lot more interest. Among these methods, the use of biometrics for identification has gradually replaced traditional knowledge-based methods and become the mainstream. For example, fingerprint has already occupied a vast market in the smartphone authentication system. Novel in-display fingerprint technique can directly install sensor under display screen, so no additional space is required to place the sensor, which is more convenient to use and provides a solution to perform fingerprint authentication in full-screen smartphones. However, in-display fingerprint is easily affected by the noise of display panel and related circuits, which causes poor image quality. If we want to use traditional local-minutiae-based method, it is necessary to use high-computation and long-latency reconstruction methods to improve fingerprint quality. However, this is not friendly to edge devices that require low computational complexity. Based on the above problems, this thesis uses global orientation field to extract features by observing the angle distribution to create the frequency vector. Since the orientation field is more tolerant to noise, high-complexity reconstruction methods can be avoided. To further improve the recognition ability, we use triplet loss to train the neural network model and change the original frequency vector into a transformed vector, where it can gather self-features while pushing away other features. Finally, we use a matching-based method for fingerprint authentication. Meanwhile, a vote decision method is used to enlarge the score gap between self and other data to obtain a better result. This method doesn’t require high-complexity reconstruction processes; therefore, it is more suitable for edge devices.