Bromba Biometrics - Fingerprint Technology

Bromba
Biometrics

Our Fingerprint Technology

Home > SDKs > Our Technology

Introduction	Block Diagram	Feature Extraction
Comparison	Enrolment	Reference Update

Introduction

Our fingerprint recognition algorithms look back on a more than 10 years lasting successful evolution history. Their first great use they had around 1996, when Siemens, under the conduction of their research division, took up an old patent from 1980 which dealt with the construction of silicon-based capacitive fingerprint sensors. The first functional sensors from the Siemens research labs proved to be a real challenge to the algorithms, which have been developed by Siemens PSE in Graz (Austria). This concerned sensor size as well as image quality which was still rather poor at that time. At the same time a further fundamental patent was generated, that allowed a biometrics performance which was unrivaled by then, especially for small sensors. This patent describes the fusion of small-sized sensor images towards a larger-sized overall fingerprint image.

Our algorithms do not only perform excellently on small-sized and thus inexpensive sensors but also with large-sized and non-capacitive ones. We could prove this at the Fingerprint Verification Competition FVC 2002. There we achieved, at first go and even passing established fingerprint vendors - a considerable third place among 31 world-wide participants. Two years later, an extended version of our fingerprint algorithm gained the top position in the evaluation study "Biofinger I" of the "Bundesamt für Sicherheit in der Informationstechnik" (BSI). (Press release, BSI Biofinger study (in German))

Since 1995 our algorithms have continuously been enhanced with respect to biometric performance but also with respect to fake protection. Altogether, our algorithms and systems profit by about 50 inventions. This emphasizes the uniqueness of our technology, which in this way represents a future-proof investment for all of our SDK customers.

Block Diagram

Fingerprint authentication systems generally provide two processes: verification/identification and enrolment. Enrolment is the process of learning about the biometric features by extracting the characteristics and storing them in a database or archive.

Usually, identification and enrolment share the same sensor and the same feature extraction unit, often only using different parametrization. For example, the quality control within the Feature Extraction block may require a higher image quality during enrolment since the enrolment data determine the identification performance during normal operation.

Enrolment Mode

Sensor Device

Feature Extraction

Reference Archive

Recognition Mode

Sensor Device

Feature Extraction

Comparison & Decision

OK / NOK

Reference Archive

This block diagram need not be realized as a linear step-by-step processing. Our system also interacts with functions from other processing units to guarantee optimal results.

Feature Extraction

Preprocessing

In the first step the sensor image is normalized. This is important since image parameters may differ significantly with varying fingers, finger conditions, and sensors. During this step, local contrast enhancement, equalization of ridge/valley widths, removal of image parts which do not relate to fingerprints, local image quality determination, normalization of image resolution, and image error corrections are performed.


Raw image from sensor		Normalized image

From the image quality information, user guidance can be derived in dependence on the sensor type. Examples for a capacitive sensor are:

No finger on sensor
Sensor coverage too small
Finger too dry
Finger too wet

The following image illustrates the negligible effect of sensor type on the binarized image which is a merit of the normalization algorithm. This way comparable images are generated even from swipe sensors, guaranteeing the use of different sensors without need for re-enrolment.

This figure demonstrates the interoperability with respect to different sensors

The preprocessing stage is also able to recognize sensor failures such as pixel, line, and column errors. These errors are detected by their gray value. If the gray value does not lie within a specified range (dependent on sensor specification), an interpolation is initiated. Only if the defective area is too large, an error message is generated.


Raw image with line and column error		Normalized image with corrected errors

In the case of joint sensors, the total sensor image is composed from several single sensors which are arranged in an array. At the edges of the different sensors images gaps may become unavoidable. In contrast to sensors failures these gaps have a fixed location. This enables additional measures for interpolation, called "bridging".


Raw image from two sensors		Binarized image without bridging		Binarized image with bridging

Minutiae Extraction

After preprocessing / normalization the minutiae are extracted. This can most reliably be performed by filtering out image noise coming from finger consistency and sensor noise. For that purpose the orientation of the ridges is determined in order to be able to filter the image exactly in the direction of the ridges. By this method the ridge noise is greatly reduced without affecting the ridge structure itself. The minutiae extraction then is done by binarizing and thinning the ridges to find the ridge endings (yellow) and bifurcations (cyan) by simple mathematical operations.


Raw image ->	Normalized image ->	Orientation field ->

Directionally filtered image ->	Binarized image ->	Thinned ridges ->

	Extracted minutiae shown as overlay

Template Generation

Our algorithm generates different types of fingerprint templates. There are two main categories of templates: large composite templates with image plus minutiae information and minutiae-only templates. For high-performance systems, image information is used in conjunction with minutiae. For standard performance where template size and processing time is essential (e.g., when the comparison runs on a smart card), only minutiae are stored.

Our composite templates may optionally be reduced in size by image compression. Since standard image compression techniques consume too much processing time in many applications, we have developed an own algorithm.

Since Triple DES encryption, as it is used for the minutiae part of the template, may be too time-consuming, we have developed an own algorithm using a symmetric key being extremely fast and robust.


Template image		Encrypted template image

Comparison

We provide three different comparison principles: the classical minutiae comparison, the correlation comparison, and the minutiae correlation comparison.

Minutiae Comparison

The Minutiae Comparison only processes elementary minutiae information such as position, type, angle, etc. A simplified version is available which can be used for low-power processing platforms like smartcards. The Minutiae Comparison is also used to deliver an initial guess for the Minutiae Correlation Comparison.

Correlation Comparison

The correlation comparison directly compares images. This comparison is used for image fusion to find common areas of different samples of the same fingerprint. Our correlation comparison compensates for translation, rotation, and even elastic deformation. Using the minutiae comparison to provide an initial guess, this comparison is used to decide over limit cases.

Minutia Correlation Comparison

The minutiae correlation comparison uses the minutiae comparison as pre-comparison. As a second component, a small image area around the minutiae is compared using correlation principles. The advantage of this procedure is a smaller template size if necessary and a much lower processing time compared with the correlation comparison without losing much accuracy. Our minutiae correlation comparison is fast enough to allow a live identification over up to 1000 references.

Enrolment

A prerequisite for authorization is enrolment, in which a biometric feature is saved as a personal reference either decentrally on a chip card or PC, or centrally in a data base. The enrolment essentially determines the biometric performance during identification and verification.

General procedure

Taking a fingerprint image which includes the features to be extracted using an appropriate fingerprint sensor
Examination of the data quality; if it is insufficient, the data are rejected and appropriate user guidance is given to improve the quality in the next trial
Extraction of the desired features from the data set and generation of a reference template
Storage of the template as "reference template" in the "reference archive"

Collection Enrolment

This enrolment method is the most simple one. It is based on the fact an identification of a sample fingerprint against a collection of, say N, samples of a fingerprint delivers much better performance if N > 1. In this case the reference template comprises N fingerprint samples which also results in a higher memory requirement. The samples may be available as minutiae lists, images, or composite templates, as described under Template Generation.

Optionally, Collection Enrolment may be combined with Fusion Enrolment. In this case, all combinations are tried to merge the fingerprints to save memory space and processing time without losing accuracy.

Comparison Score Enrolment

From a set of N enrolment fingerprints all combinations are verified against a certain set of conditions. From the combinations which fulfill all of the conditions, the best one with respect to a certain cost function is selected to deliver the reference template containing M < N fingerprints.

Optionally, Comparison Score Enrolment can be combined with Fusion Enrolment. This procedure, called Hidden Fusion tries to merge the fingerprints selected to save memory space. Those print which cannot be merged, remain unprocessed.

Fusion Enrolment

The aim of fusion enrolment is to generate a reference template from one fused fingerprint image which represent a fingerprint area larger than the sensor area. For this purpose a set of enrolment fingerprints images are merged in successive order. To be successful with a small number of trials, appropriate user guidance is provided. Fusion is done by correlation with consideration of unavoidable skin warping effects.

The advantage of this method is that it generates smaller reference templates since overlapping areas are fused and thus only appear once. Additionally, comparison time is reduced significantly. This is a fundamental precondition to use very small sensors in combination with comparison on smartcards.

Reference Update

Fusion Enrolment can be simplified significantly when online reference improvement is activated. This is a method which tries to enhance the reference template after a verification with a high degree of similarity or a high score value, respectively. This way the biometric performance (accuracy) is improved during operation by increasing size and quality of the reference template using image fusion, thus continuously reducing the FRR (False Rejection Rate).

Last update: 2010-12-14