Past Questions and Answers
Mar 6, 2017  Derek Mingyu MA  derek.ma@connect.polyu.hk 
Knowledge Structure
L12  Introduction
 User Authentication
What you know, what you have, what you are/do

History of Automatic biometrics

Biometrics Definitions
Automated methods of recognizing individuals based on their traits. A physical characteristics or personal behavioral trait used to recognize the identity, or verify the claimed identity of an enrollee
Enrollment Template Matching > matching score
Evaluation Method
 Biometrics Systems
L3  Image Processing
L46 Pattern Recognition
L79  Traditional UniModel Techs: Physical Features
L10  New Biometrics Tech
L1112  Traditional UniModel Techs: Behavioral Features
Review Points
 Iris vs Retina
 Voice feature
 Statisticsbased and Knowledgebased face detection & location
 Process: enrollment, template/feature vector and matching score
 Euclidean distances and Hamming distances
 Biometrics authentication: statistical, syntactic and NN
 Feature of signature
 PCA
 Contrast enhancement methods: power law function
 Fingerprint features, global and local
 Image processing: Convolution
 Select minimum feature vector
 Filter operations for edge extraction: Highpass filter and Laplacian edge filter
 face recognition process
Past Questions: Explanations
16A1
As you know, there are two doors used in the current EChannel border application. The first one could need to show a personal HK ID card, and the second to obtain user’s fingerprint and match with the record in the DB. If only one door to be used, how to implement this application? Please draw the necessary flowcharts to point out the common functions and the differences between these two doors and the one door EChannel border application.
Please point out some common characteristics and differences with the necessary flowchart between identification and verification.
Lec235
One door is using the identity identification. While the two door design is using the identity verification.
Identification: who are you, one to many matching, much harder, because an identification system must perform a large number of comparisons.
Verification: are you who you say you are, one to one matching
Flowchart: Lec237
More: Some systems use hierarchical or classification methods to speed up the searching. But these methods would introduce errors.
Hierarchical approach uses some simple features and fast matching algorithm to retrieve a small set of templates for further recognition using complex algorithm.
Classification approach cuts down the DB in several (fuzzy/nonfuzzy) groups. The input feature is classified to one/several group.
16A2
There are two kinds of biometrics from an eye, i.e. Iris and Retina. Please define their features and explain each advantages and disadvantages.
13C1: what kinds of features can you find from a retina image?
Retina features
Physical features
 vessels: blood transfer
 optic disk: the nerves of eye connected to brain
Pathological features
 red lesion
 bright lesion
Retina +/
Advantages
 accuracy
 stability of biometrics Sample
 resistant to fraud
 small template
Disadvantages
 difficulty to use
 consumer perceptions
 static design
 cost
 not convenient if you wear glasses or are concerned about having close contact with the reading device
Iris features
 trabecula meshwork, a tissue that gives the appearence of dividing the iris in a radial fashion
 rings
 furrows
 freckles
 corona
Iris +/
Advantages
 high level of accuracy
 unique structure for each iris
 capable of reliable identification as well as verification
Disadvantages
 potentially low contrast pattern in dark irises
 some user don’t accept eyebased technology
 high cost capture devices or inconvenient devices
 not easy to use since light sensitivity of humans
 accuracy decreases when users wear eyeglass, obscured by eyelashes, lenses/reflections
 any unusual lighting situations may affect the ability of the camera to acquire its subject
13A1, 12A1
Please compare the features between two kinds of biometrics: Back Vein and Palmprint, and then explain which one is more accurate, why?
Please compare the features between two kinds of biometrics: Hand Shape and Palmprint, and then explain which one is more accurate, why?
Features of Palmprint: Lec1014
Geometry features: finger width; length, width, thickness and area of a palm
Texture Feature
Line Features: principal lines and wrinkles
Point Features: minutiae point; delta point; datum point
Palm Vein
16A3
The following two models are from the different speakers saying the same vowel. Please try to define some necessary features to divide these two models. What kind of features could be extracted from voice biometrics? Please list at least 3 features in detail.
Feature set: cadence, frequency, pitch and tone of an individual’s voice
Features used in real system:
 frequencyban analysis
 identification from spectrograms(energy distribution of speech signal)
 use of coarticulation(analyze the points of spectrograph where coarticulation takes place)
 formant frequencies(position of the resonances can determine the differences between the speakers)
 pitch contours(variation of the pitch during the period of utterance)
 features derived from linear prediction
16A4
Face detection & location is an important stage in face recognition. There are two main types, i.e., Statisticsbased and Knowledgebased, to implement this function. Each type could include a few methods. Could you show at least one method for each type and roughly explain how to work?
Statisticsbased method
 Subspace method
Find the subspace of face images which shown common features of faces, which is a good representation of face
 NN method
Twoclass classification: face/nonface. Need to train the NN with face and nonface images. But many kinds of nonface images which are not collected, and slow.
Knowledgebased method
 Distribution ruler of grayvaluebased
 Contour ruler
 Color information
Detect faces with the use of color information of face, as usually color of faces are different from that of background color in an image.
 Movement information
 Symmetry information
16A5
Explain following basic concepts: 1) Enrollment 2) Template/Feature Vector 3) Matching Score
Enrollment: The process by which a user’s biometrics data is initially acquired, assessed, processed and stored in the form of a template for ongoing use in a biometrics system.
Template: A mathematical representation of biometrics dataskeletonized features of a detailed image and typical values of biometrics indicators of an individual. It update over time, which can be stored in central database, mobile devices and smart cards.
Feature Vector: It frequently happens that we can measure a fixed set of $d$ features for any objects or event that we want to classify. We can think of our feature set as a feature vector $x$, where $x$ is the $d$dimensional column vector. We can also think of $x$ as being a point in a $d$dimensional feature space.
Matching scores: The matching result between two templates
16A6
There are two comparison methods in pattern recognition, Euclidean distances and Hamming distances, for decision making. What difference between these two distances? If given two words: “WHILE” and “WHORL”, what is their Hamming distance?
Lec935
Hamming distances are positive integers that represent the number of pieces of data you would have to change to convert one data point into another.
Euclidean distance is the length of the line segment that connects two coordinates.
Euclidean distance is used for numeric measures, but Hamming distance is used for discrete measures.
Range of Hamming distance is from 0 to 1. But range of Euclidean distance can be 0 to infinity.
Hamming distance: 3/5.
16A7
There are three main approaches in biometrics authentication: Statistical, Syntactic and NN. For each approach, please give a its dfn and explore a simple application.
Lec54
Statistical PR: there is an underlying and quantifiable statistical basis for the generation of patterns.
Application: Most of PR systems are based on this approach. Features are assumed generated by a state of nature or classconditioned set of probabilities and/or probability density functions.
Syntactic PR: the underlying structure of the pattern provides the information fundamental for PR.
Application: It formulate hierarchical descriptions of complex patterns built up from simpler subpatterns.
NN: neither of the above cases hold true, but we are able to develop and train an architecture to correctly associate input patterns with desired response.
Example: Lec510
14A4
There are two main functions in pattern recognition: Feature Extraction and Matching. Please explain which stage is important after all possible features are extracted. How to implement matching function?
Lec427
Not all features are useful for a special problem. Feature selection is the important stage after all possible features are extracted. It is the process of choosing input to PR system and involves judgement. It is important that the extracted features be relevant to the PR task at hand.
(Another understanding of this question, choose a stage from the two captioned. This understanding is supported by TA. Then the answer: Matching.)
How to implement matching function?
 Hypothesize a plausible solution and adjust it to fit the problem
 Create a mathematical model of the problem and derive an optimal classifier.
Pattern classification, template matching.
14A5
Please find their differences in the following three pairs of basic concepts:
Template and sample; Speech recognition and voice biometrics Textdependent speaker ID and textindependent speaker ID
Template and Sample(Lec233): system will take several samples and extract unique features from samples to create a template
Speech recognition and Voice biometrics(Lec123,9): A clear graph shows at Lec129. Speech Processing > Output/Input. Input can be voice biometrics or speech recognition. Voice biometrics is confirm people’s identities using their voice. Speech recognition is extract information from the stream of speech and figure out what the person is saying.
Textdependent speaker ID: provide utterance of key words or sentences that are the same for training and recognition. Textindependent speaker ID: verifies the identity of the individual who is speaking. The performance of verification can vary according to: the quality of the audio signal, ambient noise, the variation between enrollment and verification devices. So same device for acquisition and verification.
12A6
Please define the following basic concepts used in image processing:
Pixel Image Image histogram Point operation
Pixel: the point at which an image is sampled as known as picture elements.
Image: a spatial presentation of an object: matrix representing quantized intensity values
Image histograms: Plots of $N_j$ v.s. $j$: shows the distribution of image pixels in terms of their gray levels.
Point operations: a function is applied to every pixel in an image, which operates only on the pixel’s current value.
13A4
Generally, there exist four main stages in a given biometric system. Please indicate each function.
Lec233
 Capture: a physical or behavioral sample is captured during enrollment, identification or verification process
 Extraction: unique data is extracted from the sample and a template is created
 Comparison: the template is compared to new sample
 Match/nonmatch: system then decides if the features extracted from the new sample are a match/nonmatch.
Section B
16B1
It is necessary to perform a statistical analysis from a relative large DB… Variance of Interclass, Variance of Intraclass and Fratio
Lec1222 + Lec1320 Intraclass variability remember the difference between different classes, while interclass similarity shows the similar features of two samples. Intraclass variability can show the key characteristics of the person, and interclass similarity can help the system to remove the similar feature and leave unique characteristics.
Fratio is interclass variability/intraclass variability.
Fratio is the balanced ratio considering these two variables. The higher is the ratio, the more discriminant the feature is.
16B2
Signature features
Upper and lower envelop: the curve connecting the most up or low pixel of the signature trajectory
Vertical and horizontal projection: the count of black pixels per horizontal or vertical lines
16B3
PCA Method: list the main steps to implement this method? Point out the main advantages using this method.
Steps(Lec519)
 Divide image data into training and testing set. E.g. the training set X is composed by the first 5 samples for each class and the rest construct the test set Y.
 Express every image sample in X by a feature vector with dimension of its resolution. Calculate mean value for each group($m$).
 Calculate total scatter matrix.
 Compute the eigenvalue and eigenvectors of $S_t$, we have $\lambda_i\phi_i = S_t\phi_i, i=1,…,644$, where $\lambda_i$ and $\phi_i$ are the $i^{th}$ eigenvalue and eigenvector. Reexpress 9 eigenvectors with nonzero eigenvalues in the form of image > eigenfaces.
 Select the most principal components or eigenvectors. (ratio: selected components/total sum)
 Then the PCA projection transform $W$ is composed by $W = (\phi_1, …, \phi_n)$ (if n most principal components). Obtain transformed features sets from X and Y: $X’ = (Xm)W$; $Y’ = (Ym)W$
Advantages
 Reduce data dimensionality
 Satisfy the minimal MSE rule
 Eliminate the correlation of original data
Disadvantages
 Eigenfaces do not distinguish between shape and appearance
 PCA does not use class information
14B2
Eigenface is PCAbased method with five steps. After finishing the first four stages, we obtain 9 eigenvectors with nonzero eigenvalues in the form of image. At the fifth step, the $k$ most principal components are selected abased on the ratio $\gamma$ of the eigenvalue sum of selected components to the total sum. Please decide the value of $k$ when the threshold of $\gamma$ is 85%.
See: Lec524
13A3
What is PCA? Why can it be used for biometrics authentication?
Principal Component Analysis can reduce the number of dimensions of a data set, so the image can be further processed or visualization. It is used to calculate the vectors which best represent this small region of image space.
In biometrics, different types images should occupy different areas of the smaller region, so that we can identify a person by finding the nearest known vector in image space.
Section C
16C1
Power Law Function for contrast enhancement
$\gamma < 1$ enhance contrast in dark regions $\gamma > 1$ enhance contrast in bright regions
For first picture, $\gamma$ can be 0.5. 弧线应该是圆左上角样子
For second picture, $\gamma$ can be 3. 弧线应该是圆右下角的样子
14B12
Why Median Filter is better than LowPass Filter for noise reduction?
Lowpass filter: blurs the edge, fine detail smoothed by averaging
Median filter: fine detail passed by filter, place all pixels by neighborhood median by convolving
14C1
Contrast enhancement is an important method in the image preprocessing. How to design a transfer function(T) from input to output?
P315
16C2
Fingerprint representations can be broadly categorized into two types: global and local. Global feature characteristics includes singular points and basic ridge patterns(six classes). Local representation is based on minute details(minutiae) of finger ridges. Given the following fingerprint image, please indicate which class it is and account all each global and local feature you can find.
What class it is?
Fingerprint Classification
 Loop
 Arch
 Whorl
Global Features
 Pattern Area
 Core Point
 Type Lines
 Delta
 Ridge Count
 Basic Ridge Patterns: loop, arch, whorl
Local Features
 Ridge ending
 Bifurcation
 Dot
 Island
 Spur
 Crossover
 Bridge
 Short Ridge
More: Global representation is an overall attribute of the finger and a single representation is valid for the entire fingerprint and is typically determined by an examination of the entire finger.
A local representation consists of several components, each component typically derived from a spatially restricted region of the fingerprint.
Typically, generic representations are used for fingerprint indexing and local representations are used for fingerprint matching.
14B3
What kind of points could be shown as singular points? Could you draw three basic fingerprint classes according to singular points? How many different points could be usually indicated as fingerprint minutia? Please list each definition.
Lec738
16C3
Convolution in image processing. Compute the convolved image.
Answer for this particular question: on my notebook.
14C2
Assume that there is an IrisCode with 256 bytes by using texture feature. If 4 bits represent a feature, please compute the total number of features represented by the IrisCode
256*8/4
Section D
16D1
A 3D example for pattern extraction is defined in the following. There are four kinds of features are extracted. Please select a minimum feature vector to classify the given six objects. If each feature is a binary value (0/1), please list these objects’ representation.
Use features: edges, crosssectionsize, axis.
16D2
Fourier Function. How to explain the meaning of the Spectrum in Fig. (b) transformed by Fourier Function from Fig. (a)? For a given English letter “K”, what is the main spectrums?
Lec624
Fourier theory expresses any signal as sum of $sin$ and $cos$ function.
 The center dot is the DC(Direct Current) component, represents the average value of the image.
 The other two represent the frequency of the sine function. The one dot is just a mirrored version of the other one.
 No dots in the xdirection because the image is the same everywhere in that direction.
Draw K: Lec638
16D3
There are two typical examples of filter operations for edge extraction, which are HighPass Filter and Laplacian Edge Enhancement Filter. Both are only different in the centre weight values and their weight sums are the “1”Sum for HighPass Filter and “0”Sum mask for Edge Enhancement Filter, respectively. Could you explain what difference about their filtering results?
Highpass Filter
 The highpass filter will sharpen the whole image even for smooth parts, because the sums of weight are 1.
Laplacian Edge Enhancement Filter
 While the Laplacian Edge Enhancement Filter will only sharpen the edge and the segmentation between different smooth parts.
 It works better for edge extraction.
 The smooth parts will become darker due to the sums of its weight is 0.
14D1
100 inidividuals try to use a biometric system. There are 38 genuine individuals are accepted, 8 genuine individuals rejected, 44 imposter rejected and 10 imposter accepted. Please evaluate the biometrics system by giving the FAR and FRR. If FTE=0.5, how about ATV?
FRR = True reject/Total true FAR = False reject / Total false ATV = (1FTE)(1FRR) EER is where FAR=FRR Crossover = 1:x < x=round(1/EER)
FRR = 8/46 FAR = 10/54 ATV = 0.5*38/46 = 0.41
12D2
As an example of image denoise, could you roughly explain how to do by using function of frequency domain?
Lec620
Frequency domain relates to the Fourier transform by decomposing a function into an infinite or finite number of frequencies.
In this case, we use Fourier transform the get the spectrum. Noise is high frequency in the graph, so we remove the high frequency points in the spectrum(around the border of the spectrum) and then reverse the process to get the denoised image.