基于统计模型的现场指纹处理 (VALSE webinar, 2016.1.13) 冯建江 清华大学 自动化系 http://ivg.au.tsinghua.edu.cn/~jfeng/
内 容 1. 背景介绍 2. 现场指纹方向场估计 3. 现场指纹检测与分割
Human identification in forensics • Police often face two human identification problems: Background check Crime investigation (Who is this guy?) (Who did this?) • Biometric (especially fingerprint) is the only choice: – Criminal is not willing to tell the true name(s) – Criminal will not leave their ID cards at scenes
Two identification problems Crime investigation by matching Background check by matching unknown latent to known tenprints unknown tenprint to known tenprints Mature technology available Open problem
Crime investigation by matching Background check by matching unknown latent to known tenprints unknown tenprint to known tenprints Why latent identification is more difficult? 1. Just one finger This is obvious 2. Quality of latent is very low Why?
Capture tenprints 10 rolled fingerprints 10 plain fingerprints
Capture latent prints • Fingerprints from crime scene are called “latent” fingerprints because these fingerprints are invisible, and some special processing steps are needed to make them visible. • There are many latent processing methods. Selection of a specific method depends on residue type, surface type, age… • Powder dusting is a very common technique. 1. Dust by powder 2. Take photograph 3. Lift by tape
Rolled, plain & latent These are 3 types of prints from the same finger. • Compared to rolled and plain fingerprints, quality of latent is very low. • This poses a big challenge to feature extraction. •
Feature extraction: plain vs latent From FVC2002 DB1 Recall: 39/45 Precision: 39/39 From NIST-27 Recall: 5/12 Precision: 5/110 Same algorithm: VeriFinger Feature: minutia point
Feature extraction for latent • Because of the poor performance of algorithms, until now, feature extraction for latents has been manually performed by experts. • Manual feature extraction is time consuming & expensive. • Recently, researchers have paid increasing attention to develop more powerful feature extraction algorithms.
现场指纹识别流程 指纹 现场指纹处理 数据库 指纹 指纹 特征 特征 方向场 检测 分割 估计 提取 匹配 学术界研究重点
Orientation field (OF) estimation is critical Latent fingerprint Ridge orientation field Enhanced ridge image Minutiae
Existing methods for OF estimation • Most methods have 2 steps: – Local estimation: gradient, DFT – OF regularization: low pass filtering, global parametric model • For plain and rolled fingerprints, these algorithms work very well. Chapter 3
OF estimation for latent Gradient Manual +FOMFE Why human performs much better than the algorithm?
Even without seeing the fingerprint, we are sure that this OF must be wrong. Because we have prior knowledge on fingerprint. Fingerprint experts are good at extracting features in latent because they are very familiar with fingerprints.
Use prior knowledge • We can develop a better OF estimation algorithm by using prior knowledge of fingerprints. • How to represent, learn, and use prior knowledge? • We do not know how prior knowledge on fingerprint is represented in expert’s brain.
Basic Ideas Inspired by spelling check technique, we represent prior knowledge using dictionary. To model the statistics of fingerprints accurately, we use a set of localized dictionaries. To utilize contextual information, dictionaries are overlapping.
Represent prior knowledge via dictionary zzzzzz abxxde Ixlsoa Invalid dsfwws iuytrs yyuooj work biometric topic Valid talk add together Words Orientation patches
Error correction via dictionary beaiteful beautiful charactorestic characteristic
Why localized dictionaries 4 roughly aligned fingerprints (arch, right loop, left loop, whorl). We observed: 1. Orientation patches at the corresponding location in different fingerprints are similar; 2. Orientation patches at different locations are dissimilar. 3. Orientation patches in the center are more diverse
Why localized dictionaries The observations are also validated using statistics estimated from 398 registered orientation fields. Histogram of orientations at each location Variance of orientations at each location
Contextual information frea freak, free, flea, area? ? If we know the context, we can resolve the ambiguity. “There is no such thing free as a frea lunch”
Flowchart of OF estimation
Finger coordinate system • Premise of using localized dictionaries is defining a finger coordinate system and designing an algorithm to estimate it from fingerprints. • Origin (finger center) is geometric center of a frontal finger. axes (finger direction) is • normal to finger joint and points to fingertip. • Easy to define in photograph of finger, not easy in fingerprint.
Finger center & direction Red arrows: finger center and direction for arch, loop, and whorl fingerprints Green disk and triangle: singular points
Pose of latent
Pose estimation: face vs finger Assume face and fingerprint are upright (only center will be estimated)
Pose estimation: learning
Uncertainty about finger center • Given a prototype, we are uncertain about finger center • The uncertainty can be measured by entropy Most informative patch Least informative patch Entropy: 3.5 bits Entropy: 8.5 bits Assume fingerprint is upright
Entropy of all prototypes Entropy:
Pose estimation: finger center
Pose estimation: finger direction Analogous to detecting rotated face in an image
Pose estimation: results
Dictionary lookup • Goal: find correct orientation patch in as few as possible candidates • 3 important designs – Similarity measure – Patch size – Diversity rule
Size of localized dictionaries Image value at (�, �) is the size Image value at (�, �) is standard of localized dictionary at (�, �) deviation of orientation in training samples Larger orientation deviation corresponds to larger dictionary size.
Dictionary lookup: similarity measure • Similarity measure should be robust to severe noise. ⁄ • The similarity between two patches is defined as � � � � . � � : # orientation elements in the initial orientation patch. � � : # similar orientation elements (difference ≤ 10° ). Red indicates similar orientation elements Dictionary orientation patch Initial orientation patch similarity: 42/75
Dictionary lookup: diversity
Context-based correction: energy function After dictionary lookup, we obtain a list of candidate orientation patches Φ �,� , Φ �,� , … , Φ �,� for an initial orientation patch Θ � . We use loopy belief propagation to find a set of candidates (shown as red patches), � � � � , � � , … , � � � , which minimizes energy function �(�) .
Context-based correction: compatibility • Adjacent patches are overlapped. • Compatibility between adjacent orientation patches is measured by the similarity of orientations in the overlapping blocks. High compatibility Low compatibility
Experiments • Evaluated on NIST Special Database 27, which has 258 latents and mated rolled fingerprints. • 2 types of quantitative evaluation – Accuracy of orientation field estimation – Accuracy of matching enhanced fingerprint
OF error on NIST-27 • Data: NIST-27 and 3 subsets, good (88), bad (85), ugly (85) • Measure: average Root Mean Square Deviation (RMSD) from the manually marked orientation fields Francesco Turroni, Davide Maltoni, Raffaele Cappelli, Dario Maio: Improving Fingerprint Orientation Extraction. IEEE TIFS 2011.
OF error on FVC-onGoing FVC-onGoing FOE benchmark is an online evaluation for OF estimation algorithms. • 2 datasets: good quality (10), bad quality (50); Ground-truth OF marked by human. • Measure: RMSD between ground-truth and estimated OF •
Matching accuracy Clear gap between LocalDict and GlobalDict Clear gap between GlobalDict and 2 traditional algorithms (FOMFE, STFT) Incorporate OF algorithm into a complete fingerprint recognition system, and • evaluate the final matching accuracy. To make matching problem more realistic and challenging, 27,000 rolled fingerprints • in NIST SD14 were used as background database. Fingerprint is enhanced using OF by 6 methods. • Same minutiae extraction and matching algorithm (VeriFinger) was used. •
基于统计模型的 指纹检测、分割与增强
现场指纹分割 在字典法查询以后，检查每个位置的置信度，基于置信度图 分割：
NIST-27 实验结果
实验结果：匹配性能
对多现场指纹的检测 顺序检测，每次移除对前一个姿态投票的方向元：
实验结果：检测 对基于三种方法采集的多现场指纹库进行检测，统计检测的 召回率和查准率：
Recommend
More recommend
