• Nie Znaleziono Wyników

Fingerprints in Silicon: Towards a low-cost fingerprint detector

N/A
N/A
Protected

Academic year: 2021

Share "Fingerprints in Silicon: Towards a low-cost fingerprint detector"

Copied!
3
0
0

Pełen tekst

(1)

Maxwell11.3 July 2008

13

Fingerprints in Silicon

Towards a low-cost fingerprint detector

Several years ago the Circuit and System group started an MSc

project with the goal of finding an optimal electronic system

archi-tecture which would be capable of identifying a person based on

its fingerprint. Suppose that we can succeed to design a ‘personal

smart card’, which is activated based on fingerprint technology, then

we can forget about ‘PIN codes’ and related security problems.

Authors: Mian Qin, Michel van der Net and Rene van Leuken, TU Delft

However, many problems have to be solved. The energy requirement of the intended system is such that technology needs to advance to nano-technology. The cost of the fingerprint system needs to be reduced to less than € 1, instead of cur-rently € 100. But first of all, we need to find and design a solution that can handle the computational complexity.

An introduction to fingerprints

For over more than a century fingerprints are being used to identify persons. This is because of the many advantages it has over other biometric techniques. In the early days, person identification was based on taking an image of a finger rolled in ink. Analyses and identification was then in the hands of fingerprint experts, who could establish whether two prints matched or not.

Fingerprints are assumed to be unique for every person and remain unchanged for over a lifetime. A fingerprint is formed by

patterns of ridges and valleys. Usually lo-cal ridge characteristics are used for iden-tification, these are called minute details. Over 150 different details are identified, which are called minutiae. Only two are used in most systems and are formed by ridge endings and ridge bifurcations (Fig-ure 1).

Fingerprint identification and verification should be distinguished. Identification means that a fingerprint is compared with one or more prints from a database. When there is a match, the person is identified. Verification means that the print is com-pared with only one reference print and it is only stated whether there is a match or not. Matching is performed by comparing the minutiae found with the ones stored in a database.

Minutiae extraction and image

enhancement

Correctly extracting the minutiae infor-mation from the fingerprint image guar-antees the reliability of the system. How-ever, if the minutiae information on the image is somehow spurious or even miss-ing, how can the correct minutiae be ex-tracted? In practice, this situation is pos-sible due to unpredictable reasons such as physical damage on fingerprint, sweat, or moving fingers, etc.

In figure 2, in the blocks, new spurious minutiae may be detected or real minu-tiae may be missing by the extracting al-gorithm. While in figure 3, it is clear and easy to find each minutia just as in these blue blocks. That’s to say, the minutiae information is well preserved in the im-age.

For those images like in figure 2, it is nec-essary to do image enhancement before minutiae extraction. This may greatly restore the missing information and re-move spurious minutiae. In figure 4, the image on the left is obtained by a

Q

The SeNSoR

The fingerprint sensor we use is a capacitive based solid-state sensor. Based on varying dis-charge voltages a digital image is generated. The 156x300 sensor array yields a 500-dpi image. The advantage of using a capacitive based sensor is that instead of an optical sen-sor, imitation of a fingerprint becomes really difficult. From the image of the fingerprint, a computational engine extracts a ‘template’, which only contains a few essential parameters to describe the fingerprint.

(a) Ridge ending (b) Ridge bifurcation

Figure 1: Examples of minutiae

Hollow Scratch

Discontinuity

Figure 2: Fingerprint image in poor quality

(2)

14

Maxwell11.3 July 2008

capacity scanner, while on the right, the enhanced image shows a good improve-ment in the quality and is then ready for the further extraction.

After image enhancement, the thinning algorithm thins the ridges into lines with single pixel width, based on which minu-tia extraction algorithm can be applied to

extract the minutia details. The whole process is described in the following flow chart:

Full hardware solution

The huge computational complexity due to large numbers of spatial convolu-tion limits the applicaconvolu-tion of fingerprint systems in the embedded and real-time

domain with uniprocessor architecture. In general, parallel architectures play an important role in high speed computing. In such architecture, “Pixel parallel”, a processor array of size n is assigned to n pixels, each working for a pixel simulta-neously. While in a uniprocessor solu-tion, algorithms are carried out repeated on each pixel, with a processor array, pro-cessing be carried out on n pixels in paral-lel, resulting in an n times speedup. Since each processor works on the same algo-rithm, there is no processor idle at any time. Pixel parallel seems to be a feasible and efficient solution for us. The reason why pixel parallel solution has such a high efficiency, is that most of the algo-rithms are in parallel in pixel wise. That means every processor has the same in-structions (data flow) and works on differ-ent data streams. This is a typical SIMD (single instruction stream, multiple data stream) architecture. The basic architec-ture is shown in figure 6.

Ridge bifurcation Ridge ending

Figure 3: Fingerprint image in good quailty with highlighted minutiae

Figure 4: Enhancement effect Figure 5: Flow chart of minu-tiae extractraction algorithm

(3)

Maxwell11.3 July 2008

15

LeoN PRoCeSSoR

Gaisler Research provides the LeoN processor, a 31-bit synthesisable processor core based on the SPARC V8 architecture. The core is highly configurable, and particularly suitable for system-on-a-chip (SoC) designs.

Simulation results show that the total time cost is 0.741 second for a typical fingerprint. Similar as in software imple-mentations, the frequency estimation and Gabor filtering consume most of the time. In general, 0.741 second is surely enough for real-time application and the speed is boosted by 8.11 times compared to a software implementation!

hardware-Software co-solution

The latest developments of a System on Chip realization of the fingerprint sys-tem are based on open-source software, provided by NIST. It is called MINDTCT and forms the basis of the Automatic Fingerprint Identification System (AFIS) used by the FBI.

MINDTCT is however developed to be executed on a PC. In its basic form it can-not be used for execution on an embedded processor and associated hardware accel-erators. The original software had to be stripped to its bare minimum, which in-cluded removing many functions that are either not executed or not needed and the conversion of all floating-point operations to fixed-point versions. The functions of the software system also need further optimization. In many cases using more clever algorithms can reduce the amount of computations. Hardware accelerators are designed to speed-up the most time consuming algorithms of the software. Finding the bottlenecks is done by profil-ing the software system and findprofil-ing the most time consuming functions. In the end a synthesisable VHDL system is cre-ated that can be prototyped on a FPGA. In this solution the pre-processing,

minu-tiae extraction and post-processing steps are performed. The input is a grayscale image, the output a list of minutiae with the x, y coordinates in the image, an angle that describes the orientation and a qual-ity factor.

Minutiae extraction is one part of the whole fingerprint verification process. The second part is the matching part. NIST also provides a good performing matching system called BOZORTH3. The BOZO-RTH3 software is developed to run on a PC, so this also must be stripped down and converted to a fixed-point version. After the creation of the complete finger-print verification system, the next steps are optimizations on both the hardware and the software side. The software is far from optimal, although changes are already made; it is still not designed for execution on the LEON processor. By re-designing the complete software system

with the LEON as a target, the memory usage and execution time can be reduced by a great amount. The algorithms per-formed also need further optimization. On the hardware side, the most optimal architecture needs to be found. This in-volves a lot of testing keeping many met-rics in mind like: resource usage, execu-tion time and power usage.

Low-power fingerprint verification is cer-tainly possible, but it involves a great amount of work to be done. Both soft-ware and hardsoft-ware engineers are needed to develop a chip that will perform well. Improvements in technology increase the possibilities for the engineers. Next step: ASIC designs!

A

Figure 6: SIMD archtecture

Cytaty

Powiązane dokumenty

In given problem among the tested crossover operators it is impossible to state with no doubt which of them should be applied. The crossover operator, which in some circumstances

This weakening does not hurt us, as in the use case the backtracking path is empty, turning the second part of the disjunction to false and allowing us to draw the conclusion that

MULTI-ATTRIBUTE DBSCAN OPTIMISATION ALGORITHM FOR SHIP TRAJECTORY CLUSTERING SHIP TRAJECTORY CLUSTERING BASED ON MULTIATTRIBUTE DBSCAN OPTIMISATION ALGORITHM Currently, there are

Studies involving physical training based interventions and the possibility of different respons- es depending on intra-individual characteristics represented by ge-

With the help of Theorem 7 and the Lemma, we can easily prove Theorem 9.. , be some sequences of complex numbers. .) be an asymptotic sequence. .)... By using the previous Lemma, we

In this paper, based on the induced tree of the crossed cube in the square of a graph, a novel distributed CDS construction algorithm named CDS-ITCC-G ∗ is presented, which can

For any symmetric domain (in its bounded or unbounded) realization there exists a unique G-invariant Laplace–Beltrami operator ∆ which is defined in terms of the G- invariant

We present an approach based on the Apriori algorithm that, in a post-processing stage, will allow us to reduce decision rules obtained by means of rough sets.. We will assume that