Biometric Algorithms and Systems

Biometric Algorithms

• Biometric algorithms use automated methods for biometric systems.
• The purpose is to recognize individuals by anatomical or behavioral traits.
• The system performs automated operations to verify or identify its ownership.
• Operations include enhancement, feature extraction, quality assessment, matching and fusion, and classification/indexing.
• Storage space, and bandwidth are reduced using compression algorithms.
• Acquired biometric samples ("query") are compared with previously captured samples stored in the system database ("reference" or "gallery").
• Creating the database is called enrollment.
• Verification involves comparing "query" with biometric data of a claimed identity.
• Identification involves comparing "query" with all biometric data in the database when no identity claim is present.
• Biometric devices acquire biometric samples and produce electronic representations of high-dimensional signals (fingerprints or face images).
• The "curse of dimensionality” is avoided by extracting a more compact representation, or "template”, from the raw signal.
• Biometric algorithms are various processes used to assess and enhance biometric signal quality for comparison.
• Algorithms extract/match salient features and use information fusion.
• Compression and classification/indexing are key to optimize resources.
• Biometric techniques recognize people due to the distinct characteristics of their biometric traits.
• System accuracy can drop due to variations and noise, or intrinsic limitations of biometric sensing techniques.
• Developing robust biometric algorithms is necessary to address accuracy issues.
• The goal is to extract salient and reproducible features and match them effectively with database templates.
• Addressing these challenges requires combining techniques to achieve optimal robustness, performance, and efficiency in biometric algorithm design.

Compression

• Storage or transmission of biometric data requires compression because it can be large.
• The goal of compression is to save storage space or transmission bandwidth.
• Compression can be lossless or lossy.
• Lossless compression maintains every bit of the original signal during uncompression.
• Lossy compression achieves higher compression ratios but alters the original signal.
• Lossy compression introduce artifacts which may interfere with feature extraction and degrade matching.
• Biometric systems commonly use lossy compression, balancing data quality and representation size.
• Standardization bodies define compression protocols for each biometric type.
• Protocols allow any user to reconstruct the original signal.
• Protocols specify the compression ratio needed to preserve biometric data quality.
• Fingerprint compression standards include WSQ (500ppi) and JPEG-2,000 (1,000ppi).
• Facial image compression standards include JPEG-2,000.
• Voice data compression standards include CELP.

Quality Assessment

• Biometric quality measures the usefulness of a biometric sample based on the amount of discriminatory information.
• Quality assessment algorithms assign a quantitative score to a biometric sample.
• This is based on its character, fidelity (signal to noise ratio), or utility (correlation with system performance).
• Quality measures can provide feedback upon enrollment for improvement.
• Quality measures improve operational efficiency of biometric systems.
• Local quality can assist feature extraction and assign confidence to features during matching.
• Performance of multi-biometric systems is improved through derived weights and statistical significance of individual sample or modality in fusion.
• "Bottom-up" algorithms reflect character and fidelity
• "Top-down" algorithms are based on observed utility.
• "Bottom-up" approach determines a sample’s "improvability,” or the potential to improve via recapture.
• Recapturing does not improve performance if a sample inherently lacks features.
• Recapturing may help obtain additional salient features if the signal to noise ratio is high.
• "Top-down" approach determines a performance estimate using a sample's utility.
• The estimate can disregard (emphasize) features correlated with utility.
• Biometric quality assessment algorithms and algorithms that use estimated quality information are both active research areas.
• The NIST biometric quality workshop serves as community forum for sharing research and development.
• NIST has developed/released open source software for measuring fingerprint quality.
• Standards committees work to incorporate the concept of quality into biometric standards, like ISO/IEC 29794.
• A goal is uniform interpretation and interoperability of quality scores.

Enhancement

• Enhancement improves signal quality by either reducing degradation sources or restoring signals.
• The goal is to increase the signal to noise ratio.
• Enhancement uses prior acquired signal knowledge to facilitate automatic feature extraction algorithms or improve manual processing visualization.
• Signal quality can be affected by environmental conditions, sensor noise, uncooperative subjects, and inherent low-quality biometrics.
• Additional algorithms/heuristics improve clarity of desired signal traits to ensure performance.
• Normalization (histogram equalization) or filtering (Gabor wavelets) separate noise from biometric signals.
• Segmentation, which detects the important part of the signal and discards the background, is another example of enhancement.

Feature Extraction

• Biometric data is processed to extract salient and discriminatory features that represent the underlying biometric trait during feature extraction.
• Features can be physical counterparts (fingerprint minutiae) or indirectly related to physical traits (iris image filter responses).
• The extracted set, or template, serves as input for matching and filtering (classification/indexing).
• Extracted features are ideally consistent for the same subject (small intra-class variation) and distinct between different subjects (small inter-class similarity).
• Feature extraction accuracy can be greatly affected by factors such as poor image quality and distortion.
• Feature extraction relates to dimensionality reduction by transforming the original data space into a lower dimension using most discriminatory information possible.
• The raw input signal often contains redundant and irrelevant information and is often in high dimension.
• Algorithms for standard dimensionality (PCA) are commonly employed to extract features for face images.
• The feature extraction algorithm affects matching performance, regardless of the trait.
• If feature extraction separates subjects in the feature space, simple matching algorithms can be employed.
• A matching algorithm cannot be designed if feature extraction performs poorly.
• Algorithms need to be interoperable when multiple systems work together.
• Extracted features/templates are encoded in a way that any matching system following the same encoding standard can use them.
• Common formats for storing biometric templates are defined by standardization bodies.
• This is crucial for large-scale applications, like biometric passports, where template storage space is small.
• The NIST MINEX (Minutiae Interoperability Exchange Test) has quantified the impact on system performance from the use of fingerprint minutiae standards compared to proprietary formats.

Matching

• Matching algorithms compare query-extracted features with stored database templates, producing scores for (dis)similarity between input and template.
• Variations of extracted features from modification, occlusion, presentation, and noise must be handled by matching algorithms.
• Modification includes scar, aging, or disease.
• Occlusion includes beard or glasses.
• Presentation includes pose, displacement, or nonlinear distortion.
• Noise includes lighting or motion blur.
• Presentation variations are typically handled using invariant features or by aligning templates.
• Introducing flexibility (or tolerance) in matching individual features obtains an accumulated probability value (global matching) for computing the final match score.
• This approach is shown to exhibit some complementary nature, increasing robustness to errors while preserving high accuracy.
• Integration (fusion) of various feature representations or combining different matching algorithms significantly improve matching accuracy.
• Matching must provide either a validation of claimed identity or a ranking of enrolled identification templates.
• Biometric matching algorithms range from simple nearest neighbor algorithms to sophisticated support vector machines.
• Thresholding techniques decide if the distance of the claimed identity (in verification) or first rank (in identification) is sufficient for authentication.
• In large systems, the time of an individual match must be small due to high throughput or real-time matching needs, which imposes constraints on matching algorithm design.
• Multistage matching techniques can achieve both high accuracy and speed.
• Biometric algorithms can be implemented in a parallel architecture distributing the processing of matching over CPUs.

Filtering (Classification/Indexing)

• A filtering process is employed to reduce the number of candidate hypotheses for the matching operation because one-to-one matching is computationally expensive with large-scale databases
• Filtering is achieved by classification and indexing.
• Classification algorithms, or classifiers, partition a database into a discrete set of classes.
• Classes can be defined based on global biometric data features like Henry classes for fingerprints, or implicitly derived based on data statistics.
• Biometric classification algorithms can be divided into rule-based, syntactic, structural-based, statistical, Neural Network-based and multiclassifier methods.
• Single-level classification is not efficient, as data is unevenly distributed among classes.
• More than 90% of fingerprints can belong to left loop, right loop, and whorl classes.
• Sub-classification divides some classes into more specific categories to continue narrowing down the search.
• Matching time is greatly reduced once templates in a database are classified because a query is compared only with templates from the same class.
• Indexing algorithms provide a continuous database ordering.
• This process is also referred to as continuous classification, where biometric data are no longer partitioned into disjoint classes.
• Data is associated with numerical vector representations of its main features.
• Feature vectors can be created through a similarity-preserving transformation in an extremely fast matching process.
• The matching is performed by comparing the query only with those in the database whose vector representations are close to the query in the transformed space.
• Filtering techniques often are used as a first stage in multistage matching because can be extremely fast.
• Indexing avoids classifying ambiguous data by adjusting the neighborhood size considered for matching) and can be designed as virtually error free.

Fusion

• Biometric systems can recognize a person based on information from multiple biometric sources.
• Systems that recognize people are also known as multimodal biometric systems.
• Multimodal biometric systems offer substantial improvements in enrollment and matching accuracy over unimodal systems.
• The biometric fusion algorithm combines multiple sources of information in a multimodal system.
• Fusion can be performed at the sensor, feature, match score, and decision levels.
• Fusion algorithm integrates primary biometric traits with soft biometric attributes such as fingerprint and face with gender, height, and eye color.
• Information fusion improves recognition accuracy and increases population coverage avoiding "failure to enroll" and deters spoof attacks in biometric systems.

Social Acceptance and Privacy Issues

• A biometric-based identification system's success depends heavily on human factors.
• Ease and comfort in interaction contributes to its acceptance.
• Biometric systems using face, voice, or iris may be seen as more user-friendly and hygienic given the non-touch characteristic.
• Biometric technologies needing cooperation or participation from users are more convenient
• Biometric characteristics captured without participation are a privacy threat.
• Recognition leaves trails of private information that telemarketers use to intrude.
• Privacy becomes more serious because biometric characteristics can provide additional personal information.
• Retinal patterns may provide medical information that health insurance companies use.
• There is fear that biometric identifiers could be used for linking personal information across different systems/databases.
• Biometrics ensures privacy by safeguarding identity and integrity which can be used for protecting individual privacy.
• Credit cards could be used only when the user supplies biometric characteristics (smartcard).
• Biometrics limit medical record access to authorized personnel.
• The use of personal biological characteristics in corporate or government recognition systems causes uneasiness.
• Companies and agencies that operate biometric systems should assure users that collected biometric information remains private and is used for the expressed purpose for which it was collected.
• Legislation ensures collected biometric information use remains private and misuse is punished.
• Commercial biometric systems available do not store sensed physical characteristics in their original form.
• They stored a digital representation, or template, in an encrypted format to eliminate actual physical characteristic recovery and ensure only the designated application can use the template.

Applications of Biometrics and Future Trends

• Biometric technologies are implemented in various sectors due to the increasing adoption of technology and the increasing need for security solutions
• Security and Access Control: Biometrics enhance security measures for access control systems. Biometrics provide methods of identifying/authenticating individuals, such as fingerprint, facial, and iris recognition.
• Identity Verification and Authentication: Biometrics verify individual identities across scenarios. Biometrics leverage unique physical/behavioral characteristics for reliable authentication and enhanced security and streamlines processes in banking and e-commerce.
• Time and Attendance Tracking: Biometric systems track employees' time and attendance in workplaces, schools, and other institutions for accuracy, convenience, and security. Systems use unique physical or behavioral characteristics of individuals, such as fingerprints, facial features, or iris patterns.
• Healthcare and Medical Applications: Biometrics enhance security, efficiency, and accuracy in healthcare settings. Biometrics allow for patient identification, EMR access, and identification before medication dispensing.
• Border Control and Immigration: Biometrics enhance security and streamline the travel process in border control and immigration by identifying fraudsters trying to enter/exit a country. Biometrics leverage unique physical or behavioral characteristics of individuals like fingerprints, facial features, or iris patterns.

Images and Pictures

• Human beings are visual creatures who rely on their vision to identify/classify and scan for differences/obtain feelings.
• Humans have evolved visual skills for face identification, color differentiation, and visual information processing.
• The world is in constant motion meaning even buildings/mountains change appearance based on sunlight, time of day, or shadows.
• The focus is on single images rather than changing scenes.
• An image can be a picture of a person, people, animals, an outdoor scene, a microphotograph, or from medical imaging that is not a random blur.

Image Processing

• Image processing changes image nature to either improve pictorial information or for autonomous machine perception.
• Digital image processing involves using a computer to change the nature of a digital image.

Image Acquisition and Sampling

• Sampling digits a continuous function.
• Given y = sin(x) + 1/3 sin(3x), sampling involves sampling at values of x.
• Undersampling occurs when the number of sample points is insufficient to reconstruct the function.
• All properties can be determined from a function which is sampled at 100 points is easily reconstructed.
• In order to ensure that we have enough sample points, we require that the sampling period is not greater than one-half the finest detail in our function.
• The above is known as the Nyquist criterion.
• The sampling theorem states a continuous function can be reconstructed from its samples if the sampling frequency is at least twice the maximum frequency in the function.
• Digital images are obtained by starting with a continuous scene representation.
• To view the scene, the reflected energy is recorded using visible light some other energy source.

Using Light

• Light is the predominant energy source for images; because human beings can observe it directly.
• Photographs record visual scenes.
• Digital images are captured using visible light because safe, cheap, easily detected and readily processed.
• A digital camera or a flat-bed scanner are two popular methods of producing a digital image.

Other Energy Sources

• Digital images can use energy sources other than light.
• Visible light falls in the electromagnetic spectrum that includes cosmic rays to electric power.
• Microscopy might use X-rays or electron beams.
• X-rays, having a shorter wavelength than visible light, resolve smaller objects than possible with visible light.
• X-ray tomography obtains images by encircling an object by an X-ray beam.

Applications of Image Processing

• Satellite/aerial views of land determine how much land is used for different purposes or investigate regions for crops.
• Fruit and vegetable inspection distinguishes good and fresh produce from old.
• Automatic inspection of items on a production line and inspection of paper samples.
• Fingerprint analysis, and sharpening or de-blurring of speed-camera images support law enforcement.
• Inspection of X-rays, MRIs, and CAT scans, and analysis of cell images/chromosome karyotypes support medicine.

Aspects of Image Processing

• Subdividing different image processing algorithms is convenient.
• Different algorithms exist for different tasks/problems.
• Image enhancement: Processing an image so that the result is more suitable. Examples include sharpening/de-blurring, highlighting edges, improving contrast, or removing noise.
• Image restoration: Reversing the damage done to an image by a known cause. Examples include linear motion blur, optical distortions, or periodic interference removal.
• Image segmentation: Subdividing an image into constituent parts or isolating aspects of an image. Examples include finding circles, lines, and shapes in an image and identifying cars, trees, buildings, or roads in aerial photographs.
• The classes are not disjoint so an algorithm could be used for enhancement or restoration, depending on if make it look better, or removing damage.

Image Task

• Obtaining postcodes from envelopes can be accomplished by acquiring an image using a scanner or CCD camera.
• Preprocessing is performed to render the result more suitable through contrast enhancement, noise removal, or identifying regions for the postcode.
• Segmentation extracts part of the image containing the postcode.
• Representation and descriptionextract extract particular features and corners used for differentiating digits.
• Recognition and interpretation means assigning labels to objects based on descriptors and deciphering digit strings as the postcode.

Types of Digital Images

• Four basic types of images include binary, grayscale, true color/RGB, and indexed.
• Binary images are black and white which need one bit per pixel making storage efficient and are suitable for text, fingerprints, or architectural plans.
• Greyscale images contain shades of grey from 0 (black) to 255 (white). With eight bits, grayscale images are natural for image file handling with 256 levels used in x-rays and printed works.
• True color or RGB images has a particular color, that color defined described by red, green and blue. Images contain 16,777,216 different possible colors and requires 24-bits.
• Indexed images only have a subset of the sixteen million possible colors. The image has an associated color map, or color palette, which is a list of colors used. Pixels do not give the color, but give an index to the color in the map.
• Digital image processing mostly uses sophisticated computer algorithms based on classification, feature extraction, and multi scale signal analysis pattern recognition/projection as well as with anisotropic diffusion, hidden Markov models, large editing, image restoration, independent component analysis, linear filtering, and point feature matching.
• Useful software packages include MATLAB, Image J, Amra, eFilm, Opencer and Keras.

MATLAB

• MATLAB (Matrix Laboratory) is a software package for high-level performance mathematical, computation, visualization and programing.
• MATLAB provides an interactive environment mainly for numeric computation, programming and visualization and was developed by MathWorks.
• MATLAB plots functions/data, creation of user interfaces and matrix manipulation, and supports interfacing with other programming language .
• MATLAB analyzes data, create models, and develop algorithms, and also provides built-in functions/toolboxes for mathematical operations, numerical methods and generates plots for image and signal processing.
• MATLAB is multi-paradigm because it can work with object-oriented and visual computing based on mathematical matrices and arrays.
• MATLAB is used in engineering, science, and economics with its intuitive interface and ideal computational processing of biometric data.
• Matrices are the main data type.
• MATLAB Contains a Command Window, a Command History viewer, a Workspace lister, and a Current Directory viewer.
• The Command Window is the main window where all commands are typed
• Command History displays commands across multiple sessions and can be dragged/edited into the Command Window.
• Workspace lists all variables generated, shows the type/size, and can be used to quickly plot/inspect.
• The Current Directory shows files/folders. By default, MATLAB is created in your home directory which is where to save.

Image and Signal Processing

• MATLAB offers functions for image processing tasks and commands for manipulation (reading, displaying, and saving).
• The Image Processing Toolbox offers advanced algorithms and techniques.
• MATLAB processes signals, including fingerprints, voice, and ECG.
• MATLAB functions implement operations, such as filtering, noise feature classifications.
• The Signal Processing Toolbox allows implementing digital filters, spectral analysis, and visualization.

Applications of MATLAB in Biometric Security

• MATLAB facilitates processing and analysis of biometric data in biometric security applications.
• MATLAB preprocesses biometric images or signals, extract features, and performs classification in biometric authentication systems.
• MATLAB enables researchers and practitioners to develop and evaluate algorithms/techniques for enhancing security and accuracy in biometric security.
• MATLAB serves as a platform for image/signal processing in biometric security.
• Its set of functions, user-friendly interface, is useful for researchers and practitioners, to advance the development of biometric security solutions.