371
Other Biomedical Imaging Techniques
However; voice recognition is subject to change and requires adaptive algorithms
that are much more complex than the image recognition algorithms used in fingerprint recognition, iris identification, and retinal scans.
18.9.1 BIOMETRICS METHODOLOGY
Regardless of the specific biological images used for biometrics, the process of
image registration and classification is a sensitive procedure that is rather different
from other biomedical image processing methods. Unlike tracking of the development of one single patient or comparing patients against each other for confirmation
or rejection of a diagnosis, the image processing procedure of biometrics recognition
is to process a given image and identify one of the individual whose biometric image
is previously stored in the dataset. In other words, in biometrics image processing,
the method is supposed to match the given image with all those stored in a database
and identify the best match.
Biometrics image processing often involves extracting a number of features
and characteristics from each image and matching these features with those of the
images in the database. The image processing steps involved in all of the three biometrics methods discussed involve noise reduction, image enhancement, and feature
extraction. The primary concern with both verification and authentication will be
the alignment of the respective scans, i.e., registration. The orientation will rely on
certain anatomical features in retinal scans and more on characteristics within the
image for the fingerprint analysis and the iris scan. The number of these extracted
features that are unique to each particular person range from 400 characteristics in
retinal scanning to 90 characteristics in fingerprint analysis. After these characteristics have been extracted, they are matched to the cases in the database.
The methods used for extracting the features are application specific. For instance,
in fingerprinting, a typical feature extraction method is based on the estimation of
patterns using B-splines. In other words, B-splines that are polynomial approximations are applied to describe the curvy patterns in fingerprint and then the coefficients
of these polynomials are used as the features/characteristics. The used feature can
also be based on specific subgroups of patterns observed in the fingerprint. These
features are described later in this section. As another example of feature extraction
for biometrics, consider the processing of iris images. In one particular image processing method specialized for iris matching, first the iris images are decomposed
using discrete wavelet transform (DWT) and then the DWT coefficients in particular
levels are used as the features.
The algorithms used for matching include Bayesian classification and different
families of neural networks. These algorithms are fed with the extracted image processing characteristics and are trained to identify the best match in a very large
database. Multilayer sigmoid neural networks are by far the most commonly used
algorithms for matching of biometrics.
Due to the sensitive and costly risk of misclassification in biometrics matching, a special attention is given to the error analysis. In the statistical analysis of
errors in biometrics matching, the following error concepts are widely used. A type
I error represents a false negative, which is a failure to identify the correct person.
Other Biomedical Imaging Techniques
However; voice recognition is subject to change and requires adaptive algorithms
that are much more complex than the image recognition algorithms used in fingerprint recognition, iris identification, and retinal scans.
18.9.1 BIOMETRICS METHODOLOGY
Regardless of the specific biological images used for biometrics, the process of
image registration and classification is a sensitive procedure that is rather different
from other biomedical image processing methods. Unlike tracking of the development of one single patient or comparing patients against each other for confirmation
or rejection of a diagnosis, the image processing procedure of biometrics recognition
is to process a given image and identify one of the individual whose biometric image
is previously stored in the dataset. In other words, in biometrics image processing,
the method is supposed to match the given image with all those stored in a database
and identify the best match.
Biometrics image processing often involves extracting a number of features
and characteristics from each image and matching these features with those of the
images in the database. The image processing steps involved in all of the three biometrics methods discussed involve noise reduction, image enhancement, and feature
extraction. The primary concern with both verification and authentication will be
the alignment of the respective scans, i.e., registration. The orientation will rely on
certain anatomical features in retinal scans and more on characteristics within the
image for the fingerprint analysis and the iris scan. The number of these extracted
features that are unique to each particular person range from 400 characteristics in
retinal scanning to 90 characteristics in fingerprint analysis. After these characteristics have been extracted, they are matched to the cases in the database.
The methods used for extracting the features are application specific. For instance,
in fingerprinting, a typical feature extraction method is based on the estimation of
patterns using B-splines. In other words, B-splines that are polynomial approximations are applied to describe the curvy patterns in fingerprint and then the coefficients
of these polynomials are used as the features/characteristics. The used feature can
also be based on specific subgroups of patterns observed in the fingerprint. These
features are described later in this section. As another example of feature extraction
for biometrics, consider the processing of iris images. In one particular image processing method specialized for iris matching, first the iris images are decomposed
using discrete wavelet transform (DWT) and then the DWT coefficients in particular
levels are used as the features.
The algorithms used for matching include Bayesian classification and different
families of neural networks. These algorithms are fed with the extracted image processing characteristics and are trained to identify the best match in a very large
database. Multilayer sigmoid neural networks are by far the most commonly used
algorithms for matching of biometrics.
Due to the sensitive and costly risk of misclassification in biometrics matching, a special attention is given to the error analysis. In the statistical analysis of
errors in biometrics matching, the following error concepts are widely used. A type
I error represents a false negative, which is a failure to identify the correct person.