黃漢邦臺灣大學：機械工程學研究所陳柏睿Chen, Po-JuiPo-JuiChen2007-11-282018-06-282007-11-282018-06-282004http://ntur.lib.ntu.edu.tw//handle/246246/60988大多數的指紋辨識系統，都是利用指紋的細部特徵點，作為辨識的依據。由於細微特徵比對法的辨識率，取決於指紋影像的品質，所以往往需要大量的影像前處理，來去除影像雜訊，強化指紋本體的影像。這不但會增加系統的運算負載，還會減緩辨識速度。會了避免上述缺陷，本文結合自動選取閥值法與脈差調變，配合小波轉換的特性，將指紋在空間域的訊號，轉換至頻率域中。利用指紋影像分佈在各頻帶中的能量大小，來當作指紋的特徵值，辨別不同個體的指紋，減低系統運算負載。 對於品質不佳的指紋影像，我們結合小波轉換與蓋伯函數的功能，修補指紋中破損的部分。此外，本文還利用指紋核心點與旋轉點作為輔助基準，來解決因為指紋位置的平移、旋轉所造成之頻率特徵辨識缺失的問題。從實驗結果證實，運用倒傳遞類神經網路作為辨識器，本系統最高的辨識率可達到93%以上。對於位置有所變動，或是品質不佳的指紋影像，亦能獲得此辨識率。Most fingerprint verification systems take advantage of fingerprint minutiae as matching features. Since the classification rate of the minutia-based method is determined by the quality of input fingerprint images, the minutia-based method usually demands a large amount of image preprocessing to remove signal noise. Obviously, this not only increases system computation complexity, but also reduces matching speed. In order to avoid these drawbacks, this thesis combines the automatic threshold selection with differential pulse transform algorithm and adopts the wavelet transform to transfer a fingerprint signal from the spatial domain to the frequency domain. The magnitudes of the fingerprint energies, which distribute over different frequencies, are taken as fingerprint features for identification or verification, to reduce the computation load in the system. As for low quality fingerprint images, we join the wavelet transform and Gabor filter to enhance and restore crumbling segments. In addition, we use the registration point and rotation point as auxiliary reference to solve for the problem of frequency characteristic variations induced by fingerprint translation or rotation. With back propagation neural network as classifier, experiment shows that the classification rate can achieve 93% above in our system. The same result can be obtained for the fingerprint images in varied position or with poor quality.摘要………………………………………………………………i Abstract…………………………………………………………ii List of Tables…………………………………………………v List of Figures……………………………………………….vi Chapter 1 Introduction 1 1.1 Motivation 1 1.2 Related Works 2 1.3 Objectives and Contributions 5 1.4 System Mechanism 6 1.4.1 System Fundamental Operation 7 1.4.2 ID Enrollment and ID Verification Operation 9 1.5 Thesis Organization 10 Chapter 2 Background Knowledge 12 2.1 Overview of Biometric Techniques 12 2.2 Introduction to Fingerprints 17 2.3 Fingerprint Image Acquisition Equipment 21 2.3.1 Optical Sensor 21 2.3.2 Ultrasonic Sensor 23 2.3.3 Solid-State Sensor 24 2.4 Texture Descriptions 27 2.4.1 Estimations of Dominant Local Orientation 28 2.4.2 Flow Orientation Coherence 32 2.5 Wavelet Transform 33 Chapter 3 Fingerprint Registration Point Detection 42 3.1 Image Preprocessing 43 3.1.1 Normalization 44 3.1.2 Fingerprint Enhancement 46 3.1.3 Histogram Equalization 51 3.2 Estimation of Fingerprint Orientation 52 3.3 Background Segmentation 55 3.4 Core Point Detection 57 3.5 Fingerprint Registration Point Determination 60 Chapter 4 Wavelet Features Extraction 63 4.1 Why Wavelet Features 63 4.2 Rotation Invariant Algorithm 65 4.3 Fingerprint Feature Vector 69 4.3.1 Automatic Threshold Selection 70 4.3.2 Differential Pulse Transformation 74 4.3.3 Feature Vector Estimation 75 Chapter 5 Experimental Results 79 5.1 Wavelet Feature Evaluations 79 5.1.1 KNN Classifier 80 5.1.2 BPNN Classifier 83 5.2 Rotation Test 87 5.3 System Performance 89 5.3.1 System Architecture 89 5.3.2 Test on Normal Fingerprint Images 92 5.3.3 Test on Crumbling Fingerprint Images 93 Chapter 6 Conclusions 96 6.1 Conclusions 96 6.2 Future Works 97 References 992205125 bytesapplication/pdfen-US指紋辨識指紋強化脈波差值轉換小波轉換蓋伯函數fingerprint verificationfingerprint enhancementGabor filterDifferential pulse transformwavelet transformthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/60988/1/ntu-93-R91522813-1.pdf