1. Biometric Concepts Overview¶

1.1. Introduction¶

Historically, the term biometry is related to biological sciences and refers to the application of statistical methods applied to a wide range of measurable characteristics in biology. Currently, it is more commonly used in Information Technology, consisting of electronic identification of human beings based on their physical and behavioral characteristics.

1.2. Fingerprints¶

Fingerprints are considered one of the most reliable biometric characteristics for human recognition due to their individuality and persistence. A fingerprint consists of a pattern of ridges and valleys on the surface of the fingertips and its formation is related to the earlier fetal months. Its main disadvantage is its intrusiveness since people need to cooperate explicitly when providing their fingerprints to a biometric system. Furthermore, fingerprint-based authentication is traditionally associated with criminal authentication methods. State-of-the-art authentication methods have demonstrated adequate accuracies for fingerprint recognition methods.

1.2.1. Minutiae¶

A fingerprint is characterized by a pattern of interleaved ridges (dark lines) and valleys (bright lines). Generally, ridges and valleys run in parallel and sometimes they end or bifurcate. At a global level, the fingerprint may present regions with highly sinuous patterns, called singularities. At the local level, another important feature called minutia can be found in the fingerprint patterns. Minutia (plural minutiae) derives from Latin, “small details”, and this refers to the behavior of the ridges discontinuities such as termination, bifurcation, trifurcation, or other features such as pores (small holes inside the ridges), lakes (two closed bifurcations), dots (short ridges), etc. Most systems use only terminations and bifurcations. To properly compare (match) fingerprints, the fingerprint features, such as minutiae and singularity points, must be extracted. It is also possible to extract other global information such as orientation and frequency of the ridge regions from the fingerprint. The images below show a fingerprint image captured from a fingerprint scanner and the corresponding extracted features: ridges (thin black lines) and minutiae (blue arrows for terminations and red arrows for bifurcations).

The process of computing these features from the scanned image is called extraction. It involves advanced signal processing algorithms and takes considerably more time than comparing the features of two fingerprints. Therefore, biometric systems attempt to perform it only once and store the extracted features as a dataset called fingerprint template.

1.2.2. Fingerprint Templates¶

The extracted features of a fingerprint (minutiae, ridges, singularities) are stored in a template file. These can be stored in a variety of standard and proprietary formats. Standard formats usually comprise a minimal set of information, useful for exchanging fingerprint templates between distinct biometric systems.

Proprietary fingerprint template formats usually include pre-computed properties of the minutia set that are specific to each vendor’s ABIS [1] system, allowing faster and more accurate fingerprint matching within that system.

[1]	ABIS is an acronym for Automated Biometric Identification System

1.2.2.1. ISO¶

ISO (International Organization for Standardization) and IEC (International Electrotechnical Commission) published in 2005 the ISO/IEC 19794 family of standards, which is commonly applied for biometric data formats, standardizing the common content, meaning, and representation of biometric data formats of biometric modalities, and specifies which requirements solve the complexities of applying biometrics to a wide variety of person-recognition applications, whether such applications operate in an open systems environment or consist of a single, closed system.

The ISO/IEC 19794-2 standard specifies data formats for representation of fingerprints using the notion of minutiae, defining three template formats for interchange and storage of fingerprint minutiae data. It defines a record-based format and normal and compact formats for use on a smart card. Also, it defines an extended format which can handle additional ridge counts, core, and delta location.

1.2.2.2. ANSI¶

Fingerprint images have been widely used by agencies and criminal policies around the world for criminal identification. The introduction of Automatic Fingerprint Identification Systems (AFIS) by those agencies created the demand for standard formats for interoperability and information exchange among subsystems or AFIS geographically spread.

The ANSI 378/2004 standard specifies data formats for representation of fingerprints using the fundamental notion of minutiae. The data format is generic, in that it may be applied and used in a wide range of application areas where automated fingerprint recognition is involved. No application-specific requirements or features are addressed in this standard. The Standard contains definitions of relevant terms, a description of where minutiae shall be defined, a data format for representing fingerprint data, and conformance information.

1.2.2.3. Template Consolidation¶

Two captures of the same finger are hardly ever the same, presenting small changes in minutiae positioning and quality. A few minutiae may also appear or disappear due to small changes in the way the finger was positioned over the sensor in each capture.

To increase the quality of the template, one may perform a rolled capture, but this is not always supported by the scanner device and may present its own challenges due to the rolling movement that the finger must perform. In these cases, one can perform multiple captures and combine the identified traits of the fingerprints generating a higher quality template.

The process of taking multiple captures and combining the extracted templates into one is called template consolidation. This method provides many advantages: It is able to improve the overall minutiae quality; It can discard false minutiae incorrectly detected due to capture quality problems; It is able to merge minutiae found by later captures, generating a richer template.

1.2.3. Fingerprint Imaging¶

1.2.3.1. Capture Methods¶

There are essentially 4 different capture methods for fingerprint images:

(a) Scanning paper/ink fingerprints: usually required when importing legacy records into a digital biometric system. Prints made on paper are scanned with a flatbed scanner, with resolutions ranging from 500 to 1000 dpi.

1.2.3.2. Resolution¶

Resolution is the density of pixels per area, usually measured as dots per inch (dpi). The higher the resolution of the sensor, the richer in detail the captured image is, favoring better quality templates.

For fingerprint images, a common resolution required by law enforcement agencies is 500 dpi. This means that a device with a square sensor of 1 inch by 1 inch will produce an image with approximately 250 000 pixels (500 x 500 pixels).

Note that image dimension and image resolution are distinct measurements: A device with a 2 inch by 2 inch sensor and 500 dpi resolution will produce images with dimensions 1000 x 1000, but the resolution is still 500 dpi. Both sensor area and resolution are important to ensure the quality of extracted fingerprint features.

1.2.3.3. Image Quality¶

If the fingerprint image is noisy, blurred, or contains compression artifacts, the template extractor may not be able to correctly recognize fingerprint features. For example: two ridges may become connected due to a blurred image, preventing the detection of a termination minutia.

To mitigate dust particles deposited over the sensor, some readers provide built-in noise reduction technology. Despite that, to increase the image quality it is always important to keep the capture surface clean and to follow the best practices in fingerprint capture.

To standardize fingerprint image quality evaluation, NIST (National Institute of Standards and Technology) published NFIQ in 2005. NFIQ is a standardized measure of fingerprint image quality, designed to predict performance of minutiae-based fingerprint matching systems. NFIQ rates fingerprints in 5 quality classes: NFIQ 1 (highest possible quality) to NFIQ 5 (lowest possible quality).

1.2.3.4. Image Encoding¶

In many biometric systems, it is desirable to store images for long-term purposes such as re-extracting templates with more modern extractors, or forensic and deduplication purposes.

Due to the large amount of images an ABIS usually stores, and the quality required by template extractors to efficiently perform template extraction, traditional lossy image compression algorithms are not well-suited for fingerprint image encoding, since they do not compress well enough without producing artifacts that may damage the template extraction. On the opposite side, lossless image compression algorithms generate rather large files, increasing the storage requirements.

Whenever lossless compression is used, the most common format used for fingerprints is PNG (Portable Network Graphics).

For the specific needs of an ABIS, the Wavelet Scalar Quantization format was developed.

1.2.3.4.1. Wavelet Scalar Quantization (WSQ)¶

The Wavelet Scalar Quantization format was developed by the FBI (Federal Bureau of Investigation) specifically to compress fingerprint images, reducing the compressed file size while retaining sufficient detail, especially when compared to established algorithms such as JPEG/JFIF.

The WSQ format is based on wavelet theory and was designed for compressing gray-scale fingerprint images at 500 dpi. Images with higher resolutions are better compressed with the JPEG2000 format.

1.2.4. Fingerprint Matching¶

The whole point of template extraction is to perform fingerprint comparison. Matching is the process of comparing two templates and providing a quantitative measure of similarity between them. In a working biometric system, genuine pairs (templates extracted from captures of the same finger) will yield high similarity scores, while impostor pairs (templates extracted from captures of distinct fingers) will yield low similarity scores.

Since two templates are hardly the same even if obtained from sequential captures of the same finger, matching fingerprints must take into account the changes in minutia positioning and minutiae that do not exist in either of the templates.

1.2.4.1. Score and Threshold¶

A matching algorithm has two steps. In the first step, an Evaluation Algorithm assigns a similarity score to the comparison. The second step decides if the pair is genuine or impostor using a decision threshold (DT). If the score is higher than the threshold the algorithm decides it is a genuine comparison; otherwise, it is classified as an impostor pair. The decision threshold (DT) is configurable parameter of any biometric system.

Increasing the threshold will increase false-negative errors (genuine pairs reported as impostors) and decrease the false-positive errors (impostor pairs reported as genuine); decreasing the threshold will decrease false-negative errors and increase the false-positive errors.

When choosing the decision threshold for a given application, the user must consider how acceptable false-negative and false-positive errors are, and the size of the template database.

1.3. Face Recognition¶

Face recognition is one of the most friendly and less intrusive biometric modalities. Data capture is simple and capture devices are common digital cameras, which are inexpensive.

For a biometric system, face recognition involves 3 tasks:

Detecting and delimiting faces present in an image.

Extracting a feature vector from a face, and enconding it for storage as a template.

Matching: Comparing two face templates and calculating a similarity score.

Modern face recognition algorithms are based on deep neural networks, an artificial intelligence technique where the algorithms automatically choose the best discriminant features between faces based on a large training dataset. Since the features can change as the training datasets are continuously updated, there is no standard template format for face templates. Biometric systems usually keep the original image stored so that updated templates can be extracted as training datasets improve over time.

1.4. Iris Recognition¶

The iris patterns are formed during fetal development and continue developing for the first two years of life. Iris biometry benefits from the fact that iris patterns are unique for each individual, and that they do not change after the initial development. Unlike fingerprints, which are susceptible to alteration from burns, cuts, abrasion, chemical interactions, iris patterns are very unlikely to be modified, either intentionally or accidentally. Iris capture requires cooperation from the subject, who is required to keep the eyes open and still while the capture is performed. Iris capture is performed with specialized scanners with near-infrared (NIR) lighting and NIR-sensitive sensors. The image below shows a typical iris image.

While iris capture sensors are becoming cheaper as they are integrated into consumer products such as smartphones, they are still more expensive than the cameras used to capture face images for face recognition.

For a biometric system, iris recognition involves 3 tasks:

Detecting and delimiting the pupil, the iris, and detecting any occlusions from eyelids or eyelashes.

Extracting spectral features from the iris region and encoding them for storage in a template.

Matching: Comparing two iris templates and calculation a similarity score.

The ISO 19794-6 standard, first published in 2005, defines interchange formats for iris templates. Since the actual feature extraction procedure can vary considerably between biometric system providers, the ISO formats essentially encapsulate the entire original image and annotations about the detected segmentations (pupil, iris, sclera, and a mask of occlusions). Therefore, ISO templates still require a feature extraction step before they can be compared.

1.5. Error Rates¶

A major aspect of a biometric system is its accuracy. From the user’s point of view, an error of accuracy occurs when the system fails to authenticate the identity of a registered person or when the system erroneously authenticates the identity of an intruder.

There are mainly three sources that may increase the error rate:

(a) Human Factor: Capture of biometric information always requires human participation on different degrees. For fingerprint capture, fingers must be in contact with sensor surface; on iris image capture the eyes must look at a fixed point of a digital camera. Even for face image capture, probably the less human-invasive biometric, a good capture imposes some positioning requirements of the face. As consequence, human behavior at the moment of biometric capture is a critical issue for the proceeding steps and the accuracy of recognition. An inadequate capture provokes a deficient extraction of biometrical features or even the inability for such extraction. A critical point for the assessment of biometrical systems is training people in good practices for capture of biometric records.

(b) Biometric sensors: Environmental conditions and technological properties of devices affect the quality of the captured biometric records. Also, each biometric capture technology has advantages and drawbacks.

(c) Matcher Inaccuracy: Even when biometric data is captured properly and with good quality, the matching algorithms are not perfect and may generate low similarity scores for genuine pairs and high similarity scores for impostor pairs.

1.5.1. Error Types¶

False-Positive errors happen when the biometric system classifies an impostor pair as genuine.
False-Negative errors happen when the biometric system classifies a genuine pair as an impostor.