Biometric identification: A small area fingerprint identification algorithm (III.)

Algorithm (c) aiming at the problem of small area fingerprint, this paper proposes and practices a novel method, attempts to identify a wider range of images, and effectively reduces redundant storage and computation. For fingerprints, it is possible to approximate the use of the Level3 feature, different from the algorithm (one) to the number of characteristics of the limit, also different from the algorithm (ii) will lead to hundreds of KB of template space occupancy and non-controllable computing time. This focuses on the balance of space and efficiency.

Algorithm (i) and algorithm (ii) the problem of image representation is that the image details cannot be effectively used and only some image information can be used. Similarly, it can be interpreted as general interpretation and local interpretation, and the algorithm (iii) combines the global information and local details to filter out the most relevant features of the visual concept, and removes a large amount of redundant information, which makes the algorithm dozens of times times optimized in execution time and occupy space.

The main design idea is to make use of image region invariance and local gradient, to generate a sequence with gradient information for each patch, to learn and train the sequence, and to combine the AI method to ensure its robustness.

Algorithm design is divided into two main steps:

1. Divide the image into different sliding blocks, detect valid information, and divide the results into N (n to determine the size of the description).

2. The local results are reconstructed to detect the effective information between local and local and generate Chain-code.

In order to achieve the goal more effectively, we propose an "antagonistic model", which combines the details of the image with the details, and takes a supervised learning approach to further training.

The main design of the algorithm between the algorithm (a) and the algorithm (two), that the feature will not be an isolated feature, there are "antagonistic effects" between different features, forming a close association, the Chain-code only need one-dimensional convolution to get the similarity, in order to achieve higher accuracy, the addition of AI algorithm, The robustness of the feature is better solved.

The approximate algorithm deduced from the principle is a very complicated method, and the mathematical derivation to the last is a very common formula, but it lasted a few months before the completion of the design last year, but also to take a different road.

.............................

Simple to provide a test comparison chart data, through the WB_DB8 (128x60) database test:

alg1,2,3 respectively for the algorithm mentioned in the column (i), (ii), (iii). b for the current market business algorithms ranked the forefront of the solution provider.

Summary: As of now, the market on the small area of the fingerprint identification algorithm has been introduced basically completed, but in this already very mature direction to do a little further exploration. If there are more effective solutions, we hope to learn from each other.

May later try to explore a part of the fingerprint identification related problems, such as fingerprint of the live detection, fingerprint characteristics of the security and so on, and so think of which part to add.

----------------------------------------------------------------------------------------------

2018/1/1 Update, some of the above content has been modified, thought later found that some conclusions are not accurate, the final version of the design process to the previous PPT posted out forget, too lazy to write

The algorithm was implemented using 73 people, a total of 39858 images for training, plus 13,000 noise images and 20,000 fingerprint noises generated by the model simulation (don't ask me how to accurately generate more efficient images, my heart is belly MMP).

After some of the algorithm's process and the core convolution algorithm optimization, at present can be achieved in the high-pass 8-series CPU on a single recognition of about 10ms, in high-end high-side CPU, such as high-pass 430 on a single recognition of about 40ms.

Template space is also greatly reduced, from about 120kb to 30kb or so, if added fingerprint anti-counterfeiting function about the need for 80kb of space, on the market most of the fingerprint chip can do frr<1%,far<1/20w.

Fingerprint anti-counterfeiting sample is very small, so the accuracy is not too high, if there are students have a larger false fingerprint library, welcome to communicate ...

At present, in the testing of a full screen of the screen of ultrasound and the implementation of the screen under the realization of the results will be posted some results, as well as some of the algorithm of small tricks, in addition, will update some of the evasion of fingerprint attacks.

--------------------------------------------------------------------------------------------

1. Update a set of the results of this algorithm.

Image set: 178 people, 80x64,500dpi

Comparison of the main three commercial algorithms in the market

Different from the current market for small area fingerprint algorithm scheme, this algorithm only need to offer 80kb template can be on the high-pass 4-series (430) CPU to complete the rapid identification, high-pass 6 series of millisecond speed, the majority of the market algorithm or rely on traditional machine vision algorithm, the template occupies at least 200kb, At least 10 times times slower on the same CPU!

And this algorithm only needs to register the process only 6-8 times, more than the market needs to register 12-15 times greatly shortened, but the security is indeed the market algorithm more than 10 times times!

2. Update the screen fingerprint some content and test

At present, there are three kinds of fingerprint module: optical type, capacitive type, ultrasonic type. Each has the merits and demerits, is not our article to discuss, we only talk about the algorithm ~ ~ ~

First look at the capacitive and screen ultrasound fingerprint image differences, because the screen optical fingerprint provider is reluctant to disclose data, so we only talk about the common capacitive fingerprint and screen ultrasound fingerprint, do not talk about hardware design merits:

Relatively speaking, the capacitive sensor is more mature, image imaging effect is better; screen optical sensor because of the screen package light transmittance of the current OLED screen is the only choice, the acquisition of image quality general; the screen under the ultrasonic sensor is mostly 30MHz, the theory can penetrate a variety of media, imaging quality generally, And more subtle features can be displayed.

The following is the algorithm for the screen of ultrasound fingerprint algorithm test: 28 people library, registered to use 6 images.

algo-256,512,2048, the feature dimension that is embedding for each image.

3. The story of the endpoint fingerprint algorithm attack

Although it is very dismissive of this trickery behavior, but this event also does expose the current market fingerprint algorithm part of the problem.

Link here: Burst! Phone fingerprint lock has been cracked, you dare to pay with fingerprints? ? www.sohu.com

This fingerprint solution is in fact the fingerprint of the place to paste something, but, but paste something unless the user blind eyes, or a glance can be seen, so in fact this news is a gimmick, but in response to this problem, the algorithm can be eliminated from the source, effectively avoid ....

Taking the 256-dimensional feature as an example, a continuous image is based on the characteristic autocorrelation image:

Note: The third picture above 1/3 and below 2/3 is inconsistent, that is, simulates the fingerprint false stripe.

4. Fingerprints and other

As one of the most widely used authentication methods for contact Biometric identification, fingerprint has been developed for many years, and it is possible to innovate further in the application scenario in the future. If someone asks me if my fingerprints are safe, I'll answer: unsafe. There is no absolute security, biometric identity security is inherently a probabilistic collision event, not to say far<100w, then and 6-digit password security is as high, it is inevitable that not millions will be cracked once ...

Anyway, a lot of people say I write too serious, hahaha, in fact, as an algorithm researcher, I am still Meng Da's

Reference documents:

Minutiae Extraction from Level 1 Features of fingerprint, 2016

Fingerprint indexing with pose constraint, 2016

Fingerprint matching using ridges,2006

Methodology for Partial fingerprint enrollment and authentication on Mobile Devices, 2016

aminutia-based Partial fingerprint recognition system,2005

mutual-information-based Registration of Medical Images:a Survey, 2003

Fingerprint registration by maximization of Mutual information, IEEE TIP2006

High resolution partial fingerprint alignment using Pore–valley descriptors, 2010

A New Framework for quality assessment of high-resolution fingerprint Images, IEEE Tpami 2015

A fingerprint verification System Based on triangular Matching and Dynamic time Warping,ieee Tpami 2000

A decision-theoretic generalization of OnLine learning and an application to boosting,1997

Stadnford cs229

Biometric identification: A small area fingerprint identification algorithm (III.)