Neural net Decision Support System for Hand-written Author Identifications A Ermolenko 71 78

Questions of Decision Support System on processing of hand-written identifications are considered. The author has developed an original allocation method of unique handwriting characteristics on the basis of the line and symbols analysis. The informational content criterion is modified and parametric base divisibility is shown.

 -

the introduction of coefficients  i with the formalized values of the equipment of modern facilities by alarm systems (P1), by protection of building structures (P2), by authenticity and reliability of information on the functioning of the alarm system, security alarm control codes and authorized persons of protected objects (P3) and other parameters of Pi forming:

n P   i Pi (3),

i1 R – Internal costs such as fuel costs (R1), ammunition for the training of rapid response teams (R2) and other parameters collapsible additively taking into account the coefficients by  j :

S1,S2

Thefts, fires due to non-compliance with the customer requirements for the quality of security or fire alarm, construction safety in the absence of a written refusal of the 1st person from the object

O R itry by cu re on f av rse se i

w tp ono tsn s1 ti e e a inm irequm ttonh r f oC re Z7 & OR fo tsn ,eey tn em lpo ) o ia re im iuq enm trapy r fcno tirey (ow irdh akF sceu iraevw t e r &

OR tfeh tuo tem n

b s w o a s td

y o rey lae rg p rm uh oF ap laa s

Z1 add rm

a tae is

d tim to ig ed llIe co Z2

И n o iton tc a je r b te o n e P Z3 fo lseb ss a oL lauv Z4

Given the above scenario, the l-function of the dangerous state is represented as: у  z1...z12   z1z3 z4  z2 z3 z4  z1z5  z1z6  z2 z6  z3 z4 z7  z3 z4 z8  z5 z7  z5 z8  z6 z7  z6 z8  z9  z10 z11  z10 z12

We have studied the problem links of the formalized control model. The analysis showed that the actual aspect is the prevention of decision-making on forged handwritten documents. It is important to work with handwritten documents, not only customers but also their own employees.

The presented Cause and Effect Diagram as Fishbone Diagram by Kaoru Ishikawa [ 3 ] emphasizes that in the control process there is a need for expert support of decisions by means of artificial intelligence (figure 2). These tools should be focused on the development of new methods and models of handwriting identification.

Submission of

consent to the personal data processing by one person

Z10 by ry s itecon irsvo tiireo Isnp eupS tauh Z11 &

S3 OR l saon aeg r k e a sp le tliecn taad Z12 (5)

Dishonesty of its staff An collusion with client Negligence Expertise equipment lack Limited resources for scanning and processing information Client orientation Acceptance of appeals by Fax, mail, e-mail Acceptance of consents from the 1st person for all authorized persons

Department management process Сronyism

 Internal handwritten document flow Low qualification

of employees

 Large documents array

Using intelligent methods and models to support decision-making

 Decision Support Systems The problem of decision-making when considering handwritten appeals

3. Development of Decision Support System The proposed decision support system (DSS) is based on fuzzy neural net identification (Figure 3).

Input bitmap handwritten document image s e l p m a s g n i t i r w d n a h f o e s a b a t a D e s a b a t a d r e v r e s t n e i l c

Determination of handwriting coarse (string) parameters

 Determining of main

handwriting coarse (string) parameters, including spaces and lines to highlight the word

 Normalization, equalization by weight coefficients Selection of a text fragment (words, letters) Fine (by symbols) characteristics determination Normalization, equalization by weight coefficients Decision making module by neural network Decision making module by neural network Final decision module on the basis fuzzy logic Decision making

... x12  x2

N x22  x32

... ... x1M  xNM  x2M  x3M  x2M  x4M 

...  x2M  xM 

N  ...  ... 

M M  xN 1  xN 

The analysis carried out in the course of the research showed that different characteristics of the handwriting emphasize the uniqueness of its author in different ways [ 4 ]. In order to assess the informative characteristics of handwriting and their balancing, a modification of the Shannon criterion was developed (6): t _ perj true _ maхj

 f _ false j  x  dx   f _ true j  x  dx  j  ffaallssee__ mmiaпхjj trute__pmerajхj (6)

 f _ false j  x  dx   f _ true j  x  dx false _ miпj true _ miпj

Identification within groups of handwriting parameters is realized by means of neural networks. When creating the identification module, the possibilities of using various neural networks, including self-organizing, radial-basis and probabilistic ones, were studied and tested. At the same time, the studies have shown that the optimal decision model is the back propagation network. Formation of the training sample will produce formalization of handwriting characteristics of the original image. The solution classification problem of handwriting samples is solved in the tandem "original" "falsification". Imagine the training and the test sample F_O, T_O, F_T, T_T differential form: x12  x22 x1M  x2M  x12  x32 x1M  x3M   (7) (8) this will increase the sample size to dim O  M  , where М1 – coarse (by strings) and М2 – fine (by symbols) set of parameters.

To find the optimal algorithmic complexity balance and identification quality, various methods of backpropagation neural network training (gradient method, Lewenberg-Marquardt method, quasiNewton method) has been investigated. The best convergence rate and minimize the mashing learning error EL   G  X   Y and generalization error E  G  F is been quasi-Newton BFGS algorithm

XX (Broyden — Fletcher — Goldfarb — Shannon algorithm) [ 5 ]. Algorithm functioning is based on determination gradient vector at the k-th iteration gk by multiplication Bk Pk  gk that approximates Hessian-matrix by matrix Bk for next direction optimization Pk. BFGS-conversion formula is defined as

Pk 1  gk 1  Sk  Xk1  X k – vector equal to the change of Х n the k-th iteration, Yk  gk 1  gk – the corresponding change of gradient.

In the BFGS–formula uses the results of the previous step. Thus, the 1st step should be carried out in a different way, for example, by the steepest descent method with minimizing the error of the neural network:  Sk , gk 1 Yk  Yk , gk 1  Sk   Sk , gk 1  1  Yk ,Yk  Sk (9) Yk , Sk  Yk , Sk  Yk , Sk  where y j – value of j-th output of neural network; dj – desired value of j-th output; p – number of neurons in output layer. Changing the weight will produce the formula: where h - the parameter that determines the learning speed.

The first layer of the proposed neural network topology represents 5 neurons with tangencesigmoidal activation function. The second layer consists of 1 neuron with linear activation function. The number of training epochs was 2000. Research of the identifier model have shown (Figure 4) unambiguous separability of both string and character parameters at the output of the neural network.  xmi 2  i2

E W   1 p 2

  yi  d j  , 2 j1 wij  h E wij ,

Decision-making using fuzzy logic is realized on the basis of the composite rule by Mamdani inference by means of fuzzy implication: C  z   C1'  C2'   2  C1  z   1  C2  z  . (12)

The input functions of the accessory are represented by bell-shaped distribution functions of characteristics  x  x  e . The linguistic output of the fuzzy model forms the answer about i the authenticity of the handwritten document. It is possible to apply the centroid method (center of gravity method) of output characteristic defuzzification:

n ZCT  С  Z j  Z j j1 n С  Z j  j1 4. Identification of hand-writing characteristics The coarse (string) characteristics determination of handwriting is realized by image segmentation, clustering [ 6 ], virtualization of incomplete strings clusters, two-level filtering and approximation of the reference trajectory writing of strings (figure 5). (10) (11) (13). Least squares approximation r e t l i f 1 l e v e L

а а, а а а а а а а а а а

fragment determination for fine analysis identification in the δ-corridor of a piecewise linear approximation clustering r e t l i f 2 l e v e l

After filtration, it becomes possible to approximate the reference trajectory. Consider the clustering the image. We use the filter to remove small clusters of punctuation marks and large clusters of capital letters. Cut the pair strings joined in one cluster.

The approximation is feasible by the least squares method [ 7 ]. Differentiating the residual function by unknown parameters b and c and equating derivatives to zero, we obtain a system of equations:  1 x  n  yi  bxi  c2 xi  0  122 bbx  11n  yi  bxi  c2  0 (14)

The approximation of the reference trajectory allows forming an orthogonal virtual ruler: where hav - the average clusters height,  - the parameter of the symbols deviation,  - the height of the virtual ruler.

The analysis of symbol pixel intensities when scanning with a virtual ruler allows to determine 10 coarse string parameters, as well as to select individual words for a thin (by symbol) analysis of the peculiarities of writing letters and inter-letter connections by the author (figure 6).

  hav   h2 h3     h1

The developed fine symbolic identification methods is basing on coding of the symbols contour by trapezes [ 8 ] with dynamically changing bases and their subsequent filtration.

The developed method formation of vectors begins with the selection of points pair (for example, M and the next P) of the first studied curve of the handwriting string. Having the coordinates x, y of the point, drawing a line through it Ay  Bx  C  0 with deterministic parameters A, B,C . All (15) points of contour N  xi ; yi  in the path between M , PAxi  Byi  C must meet the following condition:   const , where   (16)

Determination is realized by consecutive description of the symbol in the graph representation with formation of statistical matrix and adjacency list as well as auxiliary list (figure 7). Ayi  Bxi  C

A2  B2

. developed software is shown in figure 8.

 Time clastering Time to determine the angle of inclination

The proposed methods make it possible to formalize 14 handwriting characteristics, with their subsequent inclusion in the DSS. The evaluation of information content allows us to conclude that the most informative character characteristics are the angles of 3-beam connections, and from the number of string characteristics - the relative height of the letters and the angles of inclination. The time of operation ( t ) of the DSS consists of the work of string ( t1 ) and character ( t2 ) neural network identification, as well as generalizing fuzzy inference (  )and is not more than 7 seconds: t  t1  t2    3sec 3,5sec 0,5sec  7 sec (17)

The proposed algorithms were tested on an independent sample of 1000 samples. Confidence interval a positive identification in this case was 0.9.

The software has received state registration. The software is implemented in the client-server database concept and can be integrated into law enforcement databases. 5. Conclusion As a result of researches the decision support system allowing identifying the author of the handwritten text is synthesized. The decision model based on BFGS-neural network with the method of training on the backpropagation algorithm is selected. The developed fuzzy-neural network model of identification and methods of information processing and determination of the unique characteristics of handwriting are implemented in ready-made software that allows identifying the author of handwriting within 7 seconds of computer examination with a confidence interval of 0.9. Acknowledgments I Express my gratitude to my supervisor, doctor of technical Sciences, prof. Igor Goroshko (Management Academy of the Ministry of the Interior of the Russian Federation), my teacher Ph. D., associate Professor. Korlyakova M. O. (Bauman Moscow State Technical University) and my father's Ph. D., Assoc. Ermolenko V. A. (Bauman Moscow State Technical University) for his invaluable contribution to the development of my scientific potential.

 [1] Solojentsev E 2004 Scenario Logic and Probabilistic Management of Risk in Business and Engineering. Springer. [2] Solojentsev E 2010 Logical and Probabilistic Risk Management invalidity flight tests of machines, processes and systems // Problems of Risk Analysis, Volume 7 , N4 . pp 72 - 85 [3] Solojentsev E and Lebedev Y 2006 Logic-probabilictic models of company management unsuccess risk / Mathematical Economics, No 10, The Publishing House of the Wroclav University of Economics. Wroclav, 3 ( 40 ). [4] Vapnik V 1998 Statistical Learning Theory . NY: John Wiley [5] Avriel and Mordecai 2003 Nonlinear Programming: Analysis and Methods . - Dover Publishing [6] Castleman K 1979 Digital linage Processing. Prentice-Hall: Englewood Cliffs . - 407 p. [7] Gray , Lawrence F. Flanigan, Francis J. Kazdan , Jerry L., Frank and David H 1990 Calculus two : linear and nonlinear functions , Berlin: Springer-Verlag, p 375 [8] Bartlett P and Shawe-Taylor J 1998 Generalization performance of support vector machines and other pattern classifiers // Advances in Kernel Methods._ MIT Press, Cambridge, USA. (http://citeseer.ist.psu.edu/bartlett98generalization.html)