| 摘要: | 中文摘要：
指紋一直以來都是最為大眾所接受，也能成為法庭正式證據的生物特徵，甚至現在很多家庭或是公司的門禁管制系統，也都是採用指紋辨識模組，就是因為指紋的獨特性以及不變性，即使年紀變化或是一般受傷磨損都不會改變指紋特徵，讓指紋可以成為個人的身分表徵。
數百年前即有人開始發現指紋的特殊性，從Henry於1888提出指紋分類法之後，系統化的指紋辨識方法開始為大家所研究，各式各樣的指紋相關理論也蓬勃發展，包括指紋分類法、指紋流向、指紋特徵、指紋比對等專題也被廣泛討論。
在這本論文中，我們提出了一個完整的自動指紋辨識流程，從一開始的指紋前處理:背景去除、正規化、計算指紋流向以及頻率、奇異點的搜尋，到指紋的分類、特徵點搜尋以及比對，在這些處理流程中，我們不只參考了一些其他學者的研究，更提出了對這些流程的改善或是全新的演算法。
在指紋前處理部分，我們提出了”少即是多”的概念，就是在指紋區塊中占較小比例的圖形，反而可以決定該指紋區塊的特性，例如當指紋按壓太大力時，指紋的脊線會變得太粗、甚至會連結在一起，這時候脊線的像素會佔據大部分的區域，我們去搜尋這些脊線時會得到很多的假特徵點，因此我們反過來強化那些較少的訊號區域，就是谷線部分，反而可以得到比較正確的特徵訊息。
在指紋分類方面，我們參考資料探勘的概念，提出了一套新的指紋分類法，可以搭配Henry的分類法，也可以單獨運作。根據統計，九成五以上的指紋有上方核心點，我們利用上方核心點附近的指紋頻率變化，來進行指紋分類以及檢索，指紋會依照頻率變化被分成37類，在檢索時採用階層式架構，由機率最大的指紋分類開始比對，逐步往符合機率較低的類別進行搜尋。這個方式改善了傳統非黑即白的架構，傳統方法裡若一開始分類錯誤，就幾乎要搜尋整個資料庫來尋找可能符合的指紋。但是在我們提出的方法中，卻是由機率最高的類別逐步往機率較低的類別進行搜尋，這種階層式架構改善了指紋檢索的效率，在我們的驗證中，在第三階的分類索引上就可以達到90%以上的正確率了。
在特徵點擷取部分，我們不用傳統的二值化方法來進行搜尋，而是利用直接灰階演算的方式來進行特徵擷取，因為灰階圖形保留了較多的指紋資訊，在一些汙損或是模糊的指紋當中，進行二值化動作會產生大量的噪點以及假點，而且二值化過後，指紋圖形原始的灰階分布已經被破壞，產生錯誤時比較難以挽救，只能重新進行處理，在我們提出的方法中，我們會採用二階段搜尋，先利用灰階統計方式標出可能的脊線以及谷線，再引用” 少即是多”的概念，
由較少的圖形來強化較多的部分，最後將這些點連接之後，即可進行指紋特徵擷取。
在指紋比對部分，除了套用之前的指紋分類檢索規則之外，我們提出極座標比對法，利用前對位的方式校正指紋
`,再根據其特徵點分布進行比對，這種採用前對位的比對方式非常快速，我們可以在一秒內比對超過一萬個指紋，而同時保有相當的準確率。
最後我們會討論一些模糊或是片斷的指紋上所遭遇的問題以及未來可能的發展方向，並且提供我們進行研究階段的心得以及設計方針，提供後續研究者參考。
Abstract:
Fingerprint is the most popular biometric feature that is widely accepted by the public. Fingerprint is also adopted by the court to be the forensic evidence. Even now, a lot of family and company use a fingerprint recognition system as the access control system. Fingerprints possess two characteristics: Uniqueness and Invariance. The age and the damage on the finger didn’t change the pattern of fingerprint. Therefore, fingerprint can be the identity of individual.
Fingerprint had been researched for several hundred years. The first systematical classifying method of fingerprint was proposed by Henry in 1888. After Henry’s approach, many researchers dedicated on the systematic processing of fingerprint recognition, every kind of theorems and algorithms were developed to solve several fingerprint issues: fingerprint classification, fingerprint orientation fields, fingerprint feature and minutiae, and fingerprint matching.
In this dissertation, we propose a complete automatic fingerprint identification system from the pre-processing: segmentation, normalization, orientation and frequency estimation and singular points extraction. We also propose a new fingerprint classification method, and direct grey scale minutiae detection in fingerprint and the fingerprint matching. We are not only referring some researches in the literature, we also improve the original procedure and propose some new algorithm in the fingerprint processing`.
In the fingerprint pre-processing, we bring up the “Less is More” concept to the fingerprint. The significance of this concept is that pattern possess small portion in the fingerprint block would decide the characteristic of fingerprint texture. As an instance, if a finger pressed too hard on the fingerprint acquiring sensor, we will acquire a smudged fingerprint. The width of ridge lines will become too wide even can connect to other ridges. In this situation, the pixels belong to the ridge line take large portion in the fingerprint pattern. If we trace the ridge lines in the fingerprint, we will extract a lot of false minutiae. Hence, we enhance the information that takes less part of fingerprint which is the valley line in this case, and then we can extract the fingerprint features more correct.
In the fingerprint classification, we illustrate the concept form data-mining to propose a new fingerprint classifying and indexing method. Our method can co-operate with Henry’s classification or work individually. By the statistical data of fingerprints, over 95 percents of fingerprint contain an upper core. We calculate the frequency around the upper core and classify fingerprints to 37 classes. We design a hierarchical structure for fingerprint indexing form the class that has the largest probability of matching to those classes that have smaller probability. Our method improves the traditional classification that we have to search the entire database while the input fingerprint is incorrectly classified. We arrange those classes in the different indexing level from high probability to low. In our fingerprint indexing verification, we can acquire more than 90% of matching rate in the first three levels.
In the fingerprint minutiae extraction, we didn’t use traditional binarizing method to detect the minutiae, but direct extract minutiae on the grey scale image. Because the grey level image contains more information about fingerprint texture and significant amount of information may be lost during the binarization process. The binarization and thinning are time consuming; the thinning process may generate a large number of spurious minutiae. In the absence of an a priori enhancement step, most of the binarization techniques do not provide satisfactory results when applied to low-quality images. We propose two level minutiae extracting process, in the first level; we mark the peak and trough on the fingerprint histogram to indicate the position of ridge lines and valley lines. And we employ the “Less is more” concept that those textures will be enhanced by the pattern that takes fewer portions. In the second level, we link those points to extend the ridge lines and valley lines and then we can extract minutiae form the trace of ridge lines.
In the fingerprint matching process, we proposed the polar coordinate system to represent the minutiae of fingerprint and match those minutiae with a pre-alignment process. This matching algorithm is very fast that we can match more than 10000 fingerprint templates within one second. In the same time, we still can keep good accuracy of fingerprint matching.
In the last chapter, we discuss some issues about the processing of blurred or fragmental fingerprint and the future works. We also share some experience and the strategy that we design this system. |
