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

Algorithm (i) has introduced a small area fingerprint identification algorithm optional scheme, is a classic scheme, for the area is large enough and the Level2 feature is higher than the minimum limit, is a low memory occupancy, a fast implementation method. However, in some applications (such as the terminal, the requirements of small footprint, and small area means lower cost), if the collector area is further reduced, such as the acquisition of image 500dpi pixels in 100x60 (5mmx4mm), 80x80 (4mmx4mm), 60x60 ( 3mmx3mm) and so on, the solution stability of the algorithm (i) will be challenged very much. Algorithm (ii) mainly based on some traditional algorithms to explore the current market business algorithm solutions.

To the mobile phone fingerprint application scene, the sensor image size 128x60 For example, the input should be:

Only a small portion can be collected at a time, so the image used for identification and registration will have a small coverage area, and red may be the possible authentication fingerprint relative to the location of the template:

Therefore, this poses two questions:

The 1 algorithm (a) is not characteristic enough to support the decision of complete similarity.

2 There is a limited coverage area between the template images to be certified and may overlap with several presses.

Most of the applications on the terminal adopt multi-template, like the commonly used iphone5s cell phone, and the hand of Hua mi ov fingerprint recognition function, it is the template that uses 5-20 times of pressing as the finger.

As an example of the two questions raised above, we can propose a solution:

1. The first problem is the registration of small area coverage between graphs and graphs.

2. The second problem is to solve the problem of the relationship between the current image and the multi-template image.

Most of the existing commercial algorithms in the market take the traditional characteristics of old ideas (whatever name, the end of 2017), is nothing more than three points: features, models and scores. Using certain algorithm or interval sampling method to remove some step or special location as feature points, and then to the feature points around the gray-scale distribution of manual design features, such as one of the algorithm description:

This scheme can increase the accuracy of image information description to a large extent, through the use of image gradient, and then apply the traditional manual feature frame to the image matching operation. And to say that the image is particularly small, how to distinguish the uniqueness, this is a multi-template matching mechanism.

The typical performance comparison is under fmr<1/50k, the following is the approximate performance of four algorithms ():

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

1/1/2018 Update:

Program (ii) is the traditional machine vision algorithm, reference related CV Library will have different versions of the implementation, generally speaking, different implementations will have different requirements for image preprocessing, feature density, model and so on.

For example, we take an orb (Orient Rigde Binary) registration scheme, first for the selection of feature points to do the necessary image processing, algorithm (a) image processing is to obtain a complete fingerprint ridge line, so will choose the band-pass Gabor to do image filtering, And we are now in order to better get the details of the image to facilitate the calculation of gradients, and in different states like the image also has a noise distribution.

For the Algorithm (ii), generally speaking, the effect of image preprocessing has a great influence on the features and models, which directly determines whether the algorithm can be stabilized and landed. In the image processing, the method of Bayesian filtering and Deconvolution is adopted to obtain a stable image through the image normalization.

Feature selection under different noises (omit image processing)

For most terminals, the available space is limited, in order to get more compact features, and in order to reduce the time taken by the template and improve the efficiency of the algorithm, a binary value (binary) feature is adopted. A local description of each patch, the specific description algorithm can be summarized as follows:

Defines the difference between the current gradient and the individual extracts.

Two-valued feature generation: a dimension that is represented by D and to a feature

The loss function which is designed is:

which

Generally take 64 or 128-dimensional two-value features, i.e. 8bytes or 16bytes

Multi-template matching partial reference algorithm (III.).

Algorithm Performance test:

Single finger fingerprint verification time, frr about 37ms,far about 51ms, where more than the template occupies 250kb, the average multi-template occupied space 208kb

Test in image about 80*60 condition

* Other major reference algorithms involved (III)

Reference documents:

PATENT:CN 106663204 A

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