Fingerprints recognition using neural network Politecnico di Milano Polo Regionale di Como Methods and Technologies for Image Processing Author : Alessandro BAFFA 682075

Fingerprints recognition using neural network

Agenda Introduction Features of fingerprints The pattern recognition system Why using neural network? The goal of this method Preprocessing system Feature extraction and selection Invariant recognition Result References

Introduction

Features of fingerprints

The pattern recognition system

Why using neural network?

The goal of this method

Preprocessing system

Feature extraction and selection

Invariant recognition

Result

References

Features of fingerprints Fingerprints are imprints formed by friction ridges of the skin in fingers and thumbs. Their pattern are permanent and unchangeable on each finger during all the life; They are individual (the probability that two fingerprints are alike is about 1 in 1.9x10^15 ) They have long been used for identification

Fingerprints are imprints formed by friction ridges of the skin in fingers and thumbs.

Their pattern are permanent and unchangeable on each finger during all the life;

They are individual (the probability that two fingerprints are alike is about 1 in 1.9x10^15 )

They have long been used for identification

The pattern recognition system Image acquisition converting a scene into an array of numbers that can be manipulated by a computer Edge detection and thinning are parts of the preprocessing step which involves removing noise, enhancing the picture and, if necessary, segmenting the image into meaningful regions

Image acquisition converting a scene into an array of numbers that can be manipulated by a computer

Edge detection and thinning are parts of the preprocessing step which involves removing noise, enhancing the picture and, if necessary, segmenting the image into meaningful regions

Feature extraction in which the image is represented by a set of numerical “features” to remove redundancy from the data and reduce its dimension Classification where a class label is assigned to the image/object by examining its extracted features and comparing them with the class that the classifier has learned during its training stage. The main focus of this method is on these two last parts The pattern recognition system

Feature extraction in which the image is represented by a set of numerical “features” to remove redundancy from the data and reduce its dimension

Classification where a class label is assigned to the image/object by examining its extracted features and comparing them with the class that the classifier has learned during its training stage.

The main focus of this method is on these two last parts

Why using neural network? Neural network enable solutions to be found to problems where algorithmic methods are too computationally intensive or do not exist The problem of feature extraction and classification seems to be a suitable application for neural nets. They offer significant speed advantages over conventional techniques

Neural network enable solutions to be found to problems where algorithmic methods are too computationally intensive or do not exist

The problem of feature extraction and classification seems to be a suitable application for neural nets.

They offer significant speed advantages over conventional techniques

The goal of this method This proposed method is based on a data model for fingerprints that is structural rather than coordinate . This structural data model is robust with respect to traslation, rotation and distortion

This proposed method is based on a data model for fingerprints that is structural rather than coordinate .

This structural data model is robust with respect to traslation, rotation and distortion

Preprocessing system The first phase of the work is to capture the fingerprints image and convert it to a digital representation of 512x512 by 256 gray levels. Histogram equalization technique is used to increase the contrast if the illumination condition is poor But we are only interested in binary information

The first phase of the work is to capture the fingerprints image and convert it to a digital representation of 512x512 by 256 gray levels.

Histogram equalization technique is used to increase the contrast if the illumination condition is poor

But we are only interested in binary information

Preprocessing system Binarization is usually performed by using Laplacian edge detection operator Local derivative operator such as “Roberts”, “Prewitt” or “Sobel” Thresholding tecnique The binary image is further enhanced by a thinning algorithm which reduces the image ridges to a skeletal structure

Binarization is usually performed by using Laplacian edge detection operator

Local derivative operator such as “Roberts”, “Prewitt” or “Sobel”

Thresholding tecnique

The binary image is further enhanced by a thinning algorithm which reduces the image ridges to a skeletal structure

Preprocessing system The thinning algorithm while deleting unwanted points should not: Remove end points Break connectedness Cause excessive erosion of the region After obtaining the binary form of the fingerprint image, there may be some irregularities caused by skinfolds and contiguous ridges or spreading of ink due to finger pressure, and so on..

The thinning algorithm while deleting unwanted points should not:

Remove end points

Break connectedness

Cause excessive erosion of the region

After obtaining the binary form of the fingerprint image, there may be some irregularities caused by skinfolds and contiguous ridges or spreading of ink due to finger pressure, and so on..

Preprocessing system To remedy this problem, smoothing is necessary and includes: Filling holes Deleting redundant points Removing noisy points Filling potential missing points

To remedy this problem, smoothing is necessary and includes:

Filling holes

Deleting redundant points

Removing noisy points

Filling potential missing points

Feature extraction and selection Selection of good feature is a crucial step in the process since the next stage sees only these features and acts upon them. 150 different minutiae type have been identified but in practice only ridge ending and ridge bifurcation are used.

Selection of good feature is a crucial step in the process since the next stage sees only these features and acts upon them.

150 different minutiae type have been identified but in practice only ridge ending and ridge bifurcation are used.

Feature extraction and selection Good features are those satisfying two requirements: Small intraclass invariance (i.e. slightly different shapes with similar general characteristics should have numerically close values) Large interclass separation (i.e. features from different classes should be quite different numerically)

Good features are those satisfying two requirements:

Small intraclass invariance (i.e. slightly different shapes with similar general characteristics should have numerically close values)

Large interclass separation (i.e. features from different classes should be quite different numerically)

Feature extraction and selection A multilayer perceptron network of three layers is trained to detect the minutiae in the thinned part image of size 128x128 The first layer has nine units associated with the components of the input vector The hidden layer has five units The output layer has one unit corresponding to the number of the classes The network is trained to output ‘1’ when the input window is centered on the feature to be located and it outputs ‘0’ if minutiae are not present

A multilayer perceptron network of three layers is trained to detect the minutiae in the thinned part image of size 128x128

The first layer has nine units associated with the components of the input vector

The hidden layer has five units

The output layer has one unit corresponding to the number of the classes

The network is trained to output ‘1’ when the input window is centered on the feature to be located and it outputs ‘0’ if minutiae are not present

Feature extraction and selection the network is trained by using the backpropagation learning technique and the weight change is updated according to

the network is trained by using the backpropagation learning technique and the weight change is updated according to

Feature extraction and selection The trained network is then used to analyze the complete image by raster scanning the fingerprint via window of size 3x3 In order to prevent the falsely reported features and select “significant” minutiae, two more rules are added to the system to guarantee perfect ridge forks are detected while excluding all other features: At those potential minutiae feature points we examine them by increasing the window size to 5x5 If two or more minutiae are too close togheter, we ignore all of them

The trained network is then used to analyze the complete image by raster scanning the fingerprint via window of size 3x3

In order to prevent the falsely reported features and select “significant” minutiae, two more rules are added to the system to guarantee perfect ridge forks are detected while excluding all other features:

At those potential minutiae feature points we examine them by increasing the window size to 5x5

If two or more minutiae are too close togheter, we ignore all of them

Distribution of minutiae of two identical fingerprints 2(a) before and 2(b) after applying the rules Feature extraction and selection

Distribution of minutiae of two identical fingerprints 2(a) before and 2(b) after applying the rules

Invariant recognition The location of a reference point of the fingerprints is important for invariant recognition and has to be determined Contour tracing is used to find one or more turning points (i.e. points with maximum rate of change of tracing movement) This points are then used to find the reference point

The location of a reference point of the fingerprints is important for invariant recognition and has to be determined

Contour tracing is used to find one or more turning points (i.e. points with maximum rate of change of tracing movement)

This points are then used to find the reference point

Invariant recognition The Euclidean distance d(i) from each feature point i to the reference point are calculated The distance to the center confers the property of positional invariance The data are then sorted in ascending order from d(0) to d(N) this operation gives the data the property of rotational invariance In order to make the data becomes invariant to scale change, it is normalized to unity by the shortest distance d(0) , i.e. dist(i) = d(0)/d(i), i = 0..N This will weight those feature points nearer to the center more heavly because usually these points are more significant in classification.

The Euclidean distance d(i) from each feature point i to the reference point are calculated

The distance to the center confers the property of positional invariance

The data are then sorted in ascending order from d(0) to d(N)

this operation gives the data the property of rotational invariance

In order to make the data becomes invariant to scale change, it is normalized to unity by the shortest distance d(0) , i.e. dist(i) = d(0)/d(i), i = 0..N

This will weight those feature points nearer to the center more heavly because usually these points are more significant in classification.

Invariant recognition The centroidal data patterns should be shift, scale and rotational independent Also the invariant feature vectors are in the range [0,1] and they can be directly used as the training/stored vectors in the MLP classifier

The centroidal data patterns should be shift, scale and rotational independent

Also the invariant feature vectors are in the range [0,1] and they can be directly used as the training/stored vectors in the MLP classifier

Result The recognition rate of fingerprints depends much on the quality of the fingerprints and effectiveness of the preprocessing system Such as the thresholding level used in edge detection If there are too many broken lines or noisy points in the image, the preprocessing system contour tracing may fail. An intelligent connection algorithm to recover broken lines and suppress spurious irregularities is necessary

The recognition rate of fingerprints depends much on the quality of the fingerprints and effectiveness of the preprocessing system

Such as the thresholding level used in edge detection

If there are too many broken lines or noisy points in the image, the preprocessing system contour tracing may fail.

An intelligent connection algorithm to recover broken lines and suppress spurious irregularities is necessary

References W.F. Leung – S.H. Leung, W.H. Lau – Andrew Luk, Fingerprints recognition using neural network M.T. Leung – W.E. Engeler – P. Frank, Fingerprints image processing using neural network Jacques de Villiers – Etienne Barnard, Backpropagation neural nets with one and two hidden layers Andrew Luk – S.H. Leung – C.K. Lee – W.H. Lau, A two level classifier for fingerprint recognition

W.F. Leung – S.H. Leung, W.H. Lau – Andrew Luk, Fingerprints recognition using neural network

M.T. Leung – W.E. Engeler – P. Frank, Fingerprints image processing using neural network

Jacques de Villiers – Etienne Barnard, Backpropagation neural nets with one and two hidden layers

Andrew Luk – S.H. Leung – C.K. Lee – W.H. Lau, A two level classifier for fingerprint recognition