Development the algorithm of automatic searching man's faces defined points by photo

development the algorithm of automatic searching man's faces defined points by photo

Dmitriy L. Semenov The Scientific head - Professor  Sergey I. Malafeev

Department of Information Technologies Vladimir  State University Gorky st. 87, Vladimir, 600000, RUSSIA Tel: +7(4922) 279992, E-mail:  semenof@kovrov.ru

Abstract –  Results of development the algorithms of automatic searching human’s faces defined points on a photograph are considered. suggested algorithms differ the increased speed and accuracy.

1.  INTRODUCTION

The automatic identification of the person on a photograph is a urgent task in various areas. The basic reasons of the raised attention to this task are the presence of a wide circle of technical, economic and social systems, in which the automatic identification of the man’s person determines efficiency and safety of their functioning. Such identification is required in criminalistics, in systems of bank and information safety etc. The Special meaning has the control of a condition of the operator in man-machine systems. The developed methods can be applied to elimination of so-called "human factor", i.e. for tracking a physical and emotional condition of the operator.

In clause the algorithms of the automated search defined points of the man’s person in a photo for procedure of identification are considered. The existing similar algorithms characterize by low efficiency, since allow to find the limited quantity of collected points at the large time of processing of an initial photo. The purpose of development of algorithm is the increase of its speed and accuracy.

2.  algorithm of searching anthropometric points

2.1.  Face area searching algorithm

Generally, it is supposed, that on the image except for the man’s face, there are also extraneous objects (background). For effective job of algorithm, it is necessary to remove them. In algorithm it is carried out with using the Sobel operator [1]. The convolution of the initial image with the masks is applied for this purpose

;.                                                                                                          

2.2.   Image binarization

The received image with the selected area of the person is binarized, i.e. the color image will be transformed in monochrome. The most effective method of global binarization, both on quality of processing, and on time is the Ots method [2]. The method uses histogram of distribution of meanings of pixel's brightness of the raster image. Histogram constructs on meanings of , где N – total amount of pixels on the image; n_i – amount of pixels with brightness level i. The range of brightness is divided into two classes with using of an optimum binarization threshold k, calculated according to discriminant criterion of the maximal interclass (object - background) dispersion of the brightness levels. To each class there corresponds relative frequencies  and :

k represents an integer from 0 up to L=255.  and  - average levels for each of two classes of the image. Further the maximal meaning of the division's quality estimation of the image on two parts is calculated:

where  - interclass dispersion   - total image dispersion.

2.3.  Pupil’s centers searching algorithm

For the received monochrome image the minima of function are searched:

where  - original image,  and  - coordinates of area of the image, on which there is a projection. The analysis of a vertical integrated projection (4) of the faces central parts allows to allocate horizontal strips of the monochrome image of width 4-5 diameters of a pupil, to which there correspond the minimal meanings of function V(x).

To the received fragments of the image the Hough transformation 3 is applied, which allows to find on the monochrome image flat curves given parametrically. For this purpose consider curves family on planes given by the parametrical equation F(a₁, a₂,…, a_n,x,y) = 0, where F – some function;  , ,…,   - parameters of curves family; x, y – coordinates on a plane. The parameters of the curves family form phase space, which each point corresponds to some curve. In view of step-type behavior of the information's representation in the computer and entrance data (image), it is required to translate continuous phase space in discrete. For this purpose the grid, splitted phase space on cells is entered, each of which corresponds to a set of curves with close meanings of parameters. Each cell of phase space can put in conformity number indicating amount of points on the image belonging even by one of a curve, appropriate to the given cell. The analysis of counters of cells allows to find on the image curves, on which the greatest amount of points lays.