Українська   Русский
DonNTU   Masters' portal

Abstract

Contents

Introduction

Interest in the processes of identification and authentication of identity in the modern world is gaining a growing practical need: from the bank card security and verification systems to forensic identification of relapsed criminals. Since one of the criteria of identity is the human's face, in an aspect of identifying a person we need procedure for dynamic facial recognition.

The problem of effective search and identification of the human face, is not trivial for the computer system and it is obvious to developers of specialized software. Working with low-contrast objects (in terms of the perception of the computer system of the human person, in contrast to the natural identification of the person to the human eye) is the most relevant for the development of artificial intelligence (AI), focuses on the systems computer (cyber) view, which in addition to the recognition procedures provide coding / decoding and storage of digital images of persons. Such systems can be viewed in aspect of classical problem with perception, and in aspect of new methods [1] of identifying an object.

1.1 Goal, object and subject of the study

The goal is to create system that will dynamically calculate and distinguish anthropometrical features, compare them with dynamically stored data for further identification of a person.

Object: the identity identification system development.

Subject: to unite face identification methods with post binary computing methods to improve the accuracy of identification.

1.2 Alleged scientific novelty

Increasing the speed and accuracy of facial recognition method by the use of post binary computing methods.

1.3 Planned results

Dynamic recognition system of anthropological facial features and their comparison with the dynamic database to identify the individual will be developed. Developing of this system includes:

  1. A search algorithm for a face in an image to further reduce the analyzed area;
  2. An algorithm for calculating of anthropometric facial features;
  3. An algorithm of comparing facial features with existing in database;
  4. Database with information about the people and data for their identification;
  5. Data structure of facial features;
  6. Post binary computing methods.

Also, the database will be created with the images to verify the quality parameters (accuracy, speed and so on.) of recognition system.

2. The overall structure of the face recognition

Regardless the great variety of face recognition algorithms we can distinguish the general structure of the process[2], which is demonstrated in the figure 1.

overall structure of the face recognition

Figure 1 – overall structure of the face recognition

The first step is the detection and localization of persons on the image (Viola-Jones method is the most effective). Monitoring involves a simplified face localization method to subsequent frames of continuous video.

At the second step the normalization of the image in the found area (geometrical and brightness, filters) takes place. The calculation and comparison of features variates between techniques and thus all reduced to a certain comparison of calculated features with the standards laid down in a data base. In this paper Elastic graph matching method is described.

2.1 Viola–Jones method

The Viola-Jones method uses the integral representation of the image - the matrix that matches the size of the original and stores the sum of all elements in each of its elements that are to the left and above current [3].

Elements presented in matrix form, and is calculated by the following formula(1):

Formula 1 (1)

, i(x', y') — the brightness of a pixel in original image.

Thus, each element of I (x, y) of the integral image contains pixel intensity sum of the rectangle (0, 0) to (x ', y').

Forming integral image takes linear time that is proportional to the number of pixels of the original image, and is carried out in a single pass. The calculation of the integral image I can be produced by the recurrence formula (2):

Formula 2 (2)

The most important advantage of the integral representation of the image is the ability to quickly calculate the amount of any rectangle of pixels (3), as well as any other shape that can be approximated by a few rectangles.

Formula 3 (3)

For a description of the desired object (face, hands, or other. Items) cascade of features is used. By itself, a cascade of Haar is a set of simple features (see Figure 2), which is considered the convolution of the image. The simplest features are used, composed of rectangles having only two levels 1 and -1. Thus, each rectangle is used in several different sizes. By convolution means s = X - Y, where Y - the sum of the pixels in a dark region, and X - the sum of pixels in the bright region.

Simple features

Figure 2 – Simple features

Such convolution designed to structure the information about the object: for example, in [6] show that, feature for a face center will always have a negative contraction (see Figure 3).

Пример получения свертки для центра лица человека (а – исходное изображение, б – наложение свертки на центр лица)

Figure 3 – Example convolution for human face center (a - original image, b - convolution of the face center ) [6]

2.2 Elastic graph matching

The essence of the method is reduced to comparison of the elastic graphs from the faces in images[2, 5]. On one of the graphs recognition stage - model - remains unchanged, while the other graph is deformed with a view to the best-fit of the model. In these systems model can be either rectangular lattice, and the structure formed by the features (anthropometric) points of the face (Fig. 4). The vertices of the graph are computed characteristic values are most often used complex values or ordered Gabor filter set (Figure 5.) - Gabor wavelet, which are calculated in a local region of a vertex of a locally by convolution with the brightness of the pixel with values of the Gabor filters (see Figure 6).

Example graph structure for facial recognition (а – regular lattice, б – graph on anthropometric features)

Figure 4 – Example graph structure for facial recognition (а – regular lattice, б – graph on anthropometric features)

The edges of the graph are weighted distances between adjacent peaks. The difference (distance discriminating characteristic) between the two graphs is calculated using a function of price the deformation that takes into account the difference between the characteristic values calculated at the vertices and edges of the degree of deformation

A set of Gabor filters

Figure 5 – A set of Gabor filters

Example of convolution face image with two Gabor filters

Figure 6 – Example of convolution face image with two Gabor filters

Deformation occurs by the displacement of each of its vertices by a distance in certain directions relative to its original location and the choice of such of its position, in which the apex of a deformable graph and the corresponding top reference graph is the minimum difference between the values of attributes (Gabor filter result). This operation is performed alternately for all graph vertices until it reaches the smallest total difference between the features and reference deformable graphs. The value of the price functions deformation at a position of a deformable graph and will be a measure of the difference between the input face image and the reference graph. This "relaxation" deformation procedure must be performed for all reference entities incorporated in the data base system. The result of the system of recognition - the standard with the best value price deformation function (see Figure 7).

Example of deformation in a regular lattice

Figure 7 – Example of deformation in a regular lattice

In some publications indicate 95-97% recognition efficiency even in the presence of different emotional expressions, and change the face angle of up to 15 °. However, a comparison of the elastic systems on graphs refer to the high computational cost of this approach. For example, to compare the face of the input image with the reference 87 spends approximately 25 seconds when operating on a parallel computer with 23 transputer [4]. In other publications on the subject time either doesn't states or states that it takes long time.

The disadvantages of this method include the high computational complexity of the recognition procedure, as well as lower process ability in memorizing the new standards and the linear dependence of the operating time on the size of the face image.

2.3 Gabor Filter

Its impulse response is defined by a sinusoidal wave (a plane wave for 2D Gabor filters) multiplied by a Gaussian function. Because of the multiplication-convolution property (Convolution theorem), the Fourier transform of a Gabor filter's impulse response is the convolution of the Fourier transform of the harmonic function and the Fourier transform of the Gaussian function[13].

The filter has a real and an imaginary component representing orthogonal directions. The two components may be formed into a complex number or used individually. Real part of the filter dsscribed by the formula (4), in imaginary part cos changes to sin.

Formula 4 (4)

, где x'=x cosθ+y sinθ; y'=-x sinθ+y cosθ. In this equation, λ represents the wavelength of the sinusoidal factor; θ represents the orientation of the normal to the parallel stripes of a Gabor function; σ is the sigma/standard deviation of the Gaussian envelope; ψ is the phase offset; γ is the spatial aspect ratio, and specifies the elasticity of the support of the Gabor function.

Conclusion

Computer vision - developing software industry, but the demand and has a large range of applications. people identification function in the photographs are actively using the software to manage your photo albums (Picasa, iPhoto and others.). By combining it with the actual parameters, you can make albums for the individual. Identification also finds application in security systems, for example, staff recognition of the facility.

In software development this project is not fully realized, however, we manage develop parts that find face and recognize some features with just two parallel Gabor wavelet. With the increasing number of filters and their parameters calibration increases the possibility of finding bigger number of features and noise cutoff. On the basis of these points we will be able to construct a graph to perform comparative analysis and identification.

The main challenges at this stage of the project is to set up the parameters of the Gabor filter and the development of algorithms for constructing graphs and its comparisons.

In future the use of means and methods of post binary computing [11, 12] to increase the accuracy of calculations. For example, some interference may be referred to as an uncertainty resulting interference will not be simply discarded but retained as potential face features.

In writing this essay master's work is not yet complete. Final completion: July 2017. The full text of work and materials on the topic can be obtained from the author or his manager after that date.

References

  1. Технология распознавания лиц / Data Систем. Товары и технологии XXI. [Электронный ресурс]. — Режим доступа: http://hardbro-ker.ru/pages/recognition. — Заглавие с экрана.
  2. Коломиец В. Анализ существующих подходов к распознаванию лиц. [Электронный ресурс] / Блог компании Синезис. — Режим доступа: http://habrahabr.ru/company/synesis/blog/238129.
  3. OpenCV шаг за шагом. Интегральное изображение. [Электронный ресурс]. — Режим доступа: http://robocraft.ru/blog/computervision/53-6.html
  4. Lades М. Distortion Invariant Object Recognition in the Dynamic Link Architecture. [Электронный ресурс] / Martin Lades, Jan C. Vorbruggen, Joachim Buhmann, Jorg Lange, Christoph v.d. Malsburg, Rolf P. Wurtz, Wolfgang Konen. — IEEE Transactions on Computers, vol. 42, No. 3, March 1993. — Режим доступа: http://www.cse.psu.edu/~rtc12/CSE597E/papers/objrecLadesMarlsberg93.pdf.
  5. Grother Р. Face Recognition Vendor Test (FRVT). Performance of Face Identification Algorithms. / Patrick Grother, Mei Ngan. — Information Access Division National Institute of Standards and Technology. — May 26, 2014 — р. 138.
  6. Мальцев А. Использование каскада Хаара для сравнения изображений [Электронный ресурс] / Мальцев Антон. — Режим доступа: https://habrahabr.ru/post/198338/. — Заглавие с экрана.
  7. Face Recognition [Электронный ресурс] / NSTC Subcommittee on Biometrics and Identity Management Room. — 7 August, 2006. — Режим доступа: http://www.biometrics.gov/documents/.
  8. Face Recognition Technology (FERET). Instructions on getting FERET database [Электронный ресурс]. / National Institute of Standards and Technology's web site. — Режим доступа: http://www.nist.gov/itl/iad/ig/feret.cfm.
  9. Fast and robust face detection and tracking: mc-jesus/face_detect_n_track. [Электронный ресурс]. — Режим доступа: https://github.com/mc-jesus/face_detect_n_track.
  10. NeoFace® Watch Generate Real-time Alerts Using Highly Accurate Face Recognition Technology. [Электронный ресурс]. / NEC Corporation of America, 2013. — Режим доступа: https://www.necam.com/docs/?id=c8a08fd5-e79c-4f00-9f37-9919318cc772.
  11. Аноприенко А. Я. Постбинарный компьютинг и интервальные вычисления в контексте кодо-логической эволюции. / А. Я. Аноприенко, С. В. Иваница — Донецк, ДонНТУ, УНИТЕХ, 2011. — 248 с
  12. Аноприенко А. Я. Тетралогика, тетравычисления и ноокомпьютинг. Исследования 2010–2012. / А. Я. Аноприенко, С. В. Иваница — Донецк: ДонНТУ, Технопарк ДонНТУ УНИТЕХ, 2012. — 308 с.
  13. Фильтр Габора [Электронный ресурс]. / Википедия – свободная энциклопедия — Режим доступа:: https://ru.wikipedia.org/wiki/Фильтр_Габора