2. Image processing. Mahmoud Mohamed Hamoud Kaid, IJECS Volume 6 Issue 2 Feb., 2017 Page No Page 20254

Transcription

1 International Journal Of Engineering And Coputer Science ISSN: Volue 6 Issue Feb. 17, Page No Index Copernicus Value (15): 58., DOI:.18535/iecs/v6i.17 Increase accuracy the recognition of obects using artificial neural network technology Mahoud Mohaed Haoud Kaid* Muawia Mohaed Ahed Mahoud AL Neelain University, Faculty of Engineering, Control Engineering Departent, Khartou, Sudan Mahoudkaid8@gail.co Abstract: This paper ais to find offers the possibility of building syste software used two techniques to identify the obects one of these techniques technique of artificial intelligence process of nerve cells called artificial neural networks applications, and other technical digital iage processing technology using a torque is changing (oents Hugh), where the syste will be able to fors of discriination Engineering regular and irregular artificial neural networks users to reduce the proportion of the isidentification of obects process and integrate it online with the identification using the torque is changing techniques. The network is trained on those fors for the first tie and then the output of the syste by giving congruence of any of these fors at high speed and accuracy. When using each technique separately there was a is ratio clearly and when erging two technologies with soe of the proportion of wrongdoing, for obects that have been training copletely absent and has to recognie these obects process eticulously. Keywords: Pictures for digital processing, for oents is variable (Hue oents), artificial neural networks 1. Introduction Sued up the way distinguish obects and recognition how onitoring systes for the iddle of the environent, and its ability to iage processing with high accuracy in order to recognie the obects and distinguish the fro the background, giving a reasoned decision about the affiliation or non- affiliation shots of saples studied[1]. There are any ways to recognie and distinguish obects in different areas, and using different techniques, but the best are the following ethods: Classic ethod, analytical, atching teplates, and ethod of artificial neural networks, these ethods do not necessarily have to be independent, but can be integrated with each other, so that the technique discriination is affected by the transition, change sie, and rotation required to distinguish the saples. Selection has occurred on the classic anner and ethod of artificial neural networks, which relies classic style ethod of oents non changing in the vector extraction features to resolve the proble of discriination three-diensional obects. Non Changing oents is offering a coplete set of descriptors for the iage, so it has a fundaental iportance in distinguishing obects theory, supports the concept of non- changing oents that deals with the properties of specific varieties of algebraic expressions that reain constant under the influence of general linear transforations. These features are obtained fro twodiensional iages, but when the process of applying this plucks soe pictures and found that it does not recognie the as has previously been training the, and over they recognie soe of the nearby shapes of iages of obects that have been training and this is a proble again, so we used artificial neural networks as a way copleentary to it, to solve these probles thus achieved two ethods with each other a high rate in recogniing pictures of obects process and found solution for the previous probles. As the researchers studied cared intense study in recent years and can process a treendous aount of inforation Parallel, at the sae tie can solve probles that can be solved using conventional algoriths also learn fro the exaples provided to the. Artificial Neural Networks used recently in digital iaging engineering and decision-aking, and proved to be effective in the areas of identification and classification of data arrangeents [3].. Iage processing Identify digital iage: as an iage f ( y, has been cut in each of the spatial coordinates and lighting levels. So it can be considered as evidence of an array deterines where the line and colun position of the point of the iage and the value of the atrix eleent specifies the aount of gray noral at that point. The eleents of such a digital atrix called picture eleents (Iage eleents) or (Pixels) or (Picture eleents) or Abbreviation (Pels) []. Sybolies the iage of a two-diensional function f ( y, of the intensity of light, referred to as the aftershock, as the value f at the sae spatial coordinates ( y, of the point gives the intensity (illuination) iage at that point. Because the light for of energy, the value of the function f ( y, ust be liited and not equal to ero, it ust eet the following relationship nuber (1): Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 54

2 DOI:.18535/iecs/v6i.17 y, (1) f The basic nature of the function f ( y, vectors: The first is the aount of light received on the scene that look at it, and the other is the aount of light reflected fro obects in the scene, it is applicable to call this the rovers illuination, and the reflectance. It referred to i ( y,, and r ( y,, respectively, as the product of these two functions is according to the relationship's nuber (), knowledge of the function: y, ix, yry, () f The picture fro the storage file is the iage of a true colors with subsequent pg are represented in the MATLAB environent to represent RGB In this representation is the color of each picture eleent cobination of colors in a single point fro three atrices are atrices of priary colors, it is converted to a gray iage in order to facilitate processing and enhancing the edges of obects to draw necessary to distinguish the fro the surrounding environent inforation []. 3. Geoetric Moent Invariants Hue (196) presented a theory of twodiensional oent invariants for planar geoetric figures based on the work of the 19th century atheaticians Boole, Cayley and Sylvester. In this theory, a set of invariants based on cobinations of regular oents using algebraic invariants were derived. These seven oents are invariant to translation, scale, and orientation. He also ipleented these oents to recognie the Latin alphabetic characters [5]. Many studies on deriving different types of invariants and using invariant Moents for obect recognition have been carried out since the introduction of Hu s Theory, Alt et. Al copared the perforance of oent invariants and Fourier descriptors in recogniing planar shapes []. 3.1 Obect Recognition by Geoetrical Moent Invariants Hue defined seven descriptors which coputed fro central oents through order three that are independent to obect translation, scale and orientation. Translation invariance is obtained by coputing oents that are noralied with respect to the center of gravity. The sie invariance can be coputed fro algebraic invariants. Fro the second and third order values of the noralied central oents, a set of seven invariant oents can be coputed which are independent of rotation.given a two-diensional iage, f (x, y), the oents of order ( p q) are defined as relationship nuber (3): p q M 1 Z N 1 Y. y p, q,1,,.., p q. f y, (3) Where: M, N - horiontal and vertical diensions of the iage. f ( y, : Intensity (gray class) at the point ( y, of the picture, we get ero torque put p, q = in the relationship (4), we get the following equation: M 1 N 1 f Z Y y, (4) The relationship represents the total values of the pixels in the iage, and naed an area of the iage. If the binary iage f ( y, then the iage area is equal to the flat area of the saple. It deterines the torque of the first class the following equations: 1 M 1 N 1 Z Y x M 1 N 1 Z Y y f f y, y, (5) Saple Center is one of the iportant baroeters to deterine the position of the saple, is the point at which owns coordinates, which is the su of x squares of which the distance of each other points within the sall body. By expressing the center of oents as follows: 1 y, 3. Central oents y Correspond to oents of inertia oents, saple.these oents variable depending on location for the center and scale draws the studied saple, and are therefore of liited use. So, it was a collection of oents to get to know is changing. This group can be derived beginning calculates the central oents given by the following equation: p q p q y p q y y f y, (6) According to the previous equation No.6 can be all of the following oents Account:,,,,,,, 1 1 1, You can write a sentence following equations: Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 55

3 DOI:.18535/iecs/v6i ,. y. y.. 1 y. 3 y. 3., 1. y. y.. y (7) Then the developent of the central oents easured as follows: pq p q p q p q ; 1 ; p q (8) 3.3 Moents Hue Non-changing These paraeters can define a set of oents non-changing as follows [5] These oents non-changing for a coplete set to describe the iage, which provides the possibility to recognie the saples with high accuracy. The torque account for a set of iages of saples trainer by the progra, and a safe in advance to be a library, the progra copares the oents calculated for the current iage coing fro the iage capture with oents stored and based on the result of the coparison progra gives decision to recognie or failure to identify the saple, which is done by calculating the distance. 3.4 Mechanis of action to recognie and discernent algorith After obtaining the previous oents, the recognition algorith and discriination copared between the vector oents of Hue to the iage of a saple with a vector of incoe plucks each iage saple stored in a coprehensive public library, using the technique discriination Euclidean distance given to the following relationship. D 1 7 () 1 Where: - Represents a line guide (Iage guide) in the library. 1 1 (9) 1 D - The distance between the vector and the iage I I vector oents incoe reference stored in the line I of the library. - Moents iage incoe vector,,,,, , 7 The distance between the vector oents calculated iage input and all of the reference vector I stored in the library and then Minial distance fro the vector and copares with a certain distance threshold is taken, and based on the result of the coparison with the decision taken was a picture of input belongs to the saples intern the or not. If the saple belongs to the obects that have been trained under the progra, the progra gives a signal that it was arrested on the target as shown in Figure (1), unless the progra is due to take a picture of new incoe and processes it. Input Iige Input Iige Input Iige Figure (1) Figure () Through the application of the progra and found that it inaccurately high percentage, we noticed the result for Hue result for Hue Figure (3) result for Hue Result Hue Result Hue Result Hue Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 56

4 DOI:.18535/iecs/v6i.17 presence of the saples did not recognie the progra although it was trained the as shown in Figure (), and ore it gives positive results for soe of the fors that have not been trained the as shown in Figure (3), and to overcoe these probles and iprove the work of the progra and increase the accuracy of the results our use of artificial neural networks, where this technique has a high ability to solve these probles and give excellent results. 4. Artificial neural network Aong countless nuber of neural network structures, there are two that are used ore often than all others: the ultilayer perceptron (MLP) and the radial basis function (RBF). These networks have been extensively studied, epirically tested on a broad spectru of applications, and hence their properties are now well known. Furtherore, these two types of networks have proven to be universal approxiates a ter referring to the ability of these networks to approxiate any decision boundary of arbitrary diensionality and arbitrary coplexity with arbitrary precision, given adequate aount of data [6]. A neuron with a single R-eleent input vector is shown below. Here the individual eleent inputs p 1, p,... p R are ultiplied by weights w 1, 1, w 1,,... w 1, R and the weighted values are fed to the suing unction. Their su is siply Wp, the dot product of the (single row) atrix W and the vector p. Figure (5) The neuron has a bias b, which is sued with the weighted inputs to for the net input n. This su, n, is the arguent of the transfer function f [7]. 4.4 Training a Network Conceptually, a network forward propagates activation to produce an output and it backward propagates error to deterine weight changes (as shown in Figure 6). The weights on the connections between neurons ediate the passed values in both directions. Figure (4) Scheatic illustration of the neuron 4.1 Neuronal Model and the Multilayer Perceptron. The artificial neural networks ANN or siply neural networks are designed to iic the decisionaking ability of the brain, and hence their structure resebles that of the nervous syste, albeit very crudely, featuring a assively interconnected set of eleents. In its ost basic for, a single eleent of the nervous syste, a neuron, consists of four basic parts, as scheatically illustrated in Figure(4): the dendrites that receive inforation fro other neurons, the cell body (the soa) that contains the nucleus and processes the inforation received by the neuron, the axon that is used to transit the processed inforation away fro the soa and toward its intended destination [6]. 4. Transfer Functions Many transfer functions are used in neural network. All of the atheatical transfer functions can be realied with a function having the sae nae. The sigoid transfer function shown below takes the input, which ay have any value between plus and inus infinity, and squashes the output into the range to 1. This transfer function is coonly used in backpropagation networks, in part because it is differentiable [8]. 4.3 Neuron with Vector Input Figure (6) The Backpropagation algorith is used to learn the weights of a ultilayer neural network with a fixed architecture. It perfors gradient descent to try to iniie the su squared error between the network s output values and the given target values. Figure (7) depicts the network coponents which affect a particular weight change. Notice that all the necessary coponents are locally related to the weight being updated. This is one feature of backpropagation that sees biologically plausible. However, brain connections appear to be unidirectional and not bidirectional as would be required to ipleent backpropagation. The change to a hidden to output weight depends on error (depicted as a lined pattern) at the output node. While the change to a input to hidden weight depends on error at the hidden node (which in turn depends on error at all the output nodes) and activation at the input node. Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 57

5 DOI:.18535/iecs/v6i.17 Note that only one ter of the net suation will have a non-ero derivative: again the one associated with the particular weight we are consider in Weight change rule for a hidden to output weight Now substituting these results back into our original equation we have: Notation For the purpose of this derivation, we will use the following notation: The subscript k denotes the output layer. The subscript denotes the hidden layer. The subscript i denotes the input layer. w k: denotes a weight fro the hidden to the output layer. w i: denotes a weight fro the input to the hidden layer. a: denotes an activation value. t: denotes a target value. net: denotes the net input Gradient Descent on Error We can otivate the backpropagation learning algorith as gradient descent on su-squared error (we square the error because we are interested in its agnitude, not its sign). The total error in a network is given by the following equation () (the 1/ will siplify things later). We want to adust the network s weights to reduce this overall error [4]. We will begin at the output layer with a particular weight. However error is not directly a function of a weight. We expand this as follows [9]. Let s consider each of these partial derivatives in turn. Derivative of the error with respect to the activation is given by the equation a ter in the forula for, the error signal : We d like to be able to rewrite this result in ters of the activation function. Notice that Using this fact, we can rewrite the result of the partial derivative as: Derivative of the net input with respect to a weight Notice that this looks very siilar to the Perceptron Training Rule. The only difference is the inclusion of the derivative of the activation function. This equation is typically siplified as shown below where the ter repesents the product of the error with the derivative of the activation function Weight change rule for an input to hidden weight Now we have to deterine the appropriate weight change for an input to hidden weight. This is ore coplicated because it depends on the error at all of the nodes this weighted connection can lead [9]. ( ) ( ) 5. Create a feed-forward backpropagation network with Matlap Purpose 1create a feed-forward backpropagation network Syntax net = newff net = newff (PR,[S1 S...SNl],{TF1 TF...TFNl},BTF,BLF,PF) Description net = newff creates a new network with a dialog box. Newff (PR,[S1 S...SNl],{TF1 TF...TFNl},BTF,BLF,PF) takes, PR R x atrix of in and ax values for R input eleents Si Sie of ith layer, for Nl layers TFi Transfer function of ith layer, default = 'tansig' BTF Backpropagation network training function, default = 'trainl' BLF Backpropagation weight/bias learning function, default = 'learngd' PF Perforance function, default = 'se' The transfer functions TFi can be any differentiable transfer function such as tansig, logsig, or purelin [7]. In either case, calling train with the resulting network will train the network with trainl. 5.1 trainl Purpose1Levenberg-Marquardt backpropagation Syntax[net,TR]= trainl(net,pd,tl,ai,q,ts,vv,tv) info = trainl(code) Description trainl is a network training function that updates weight and bias values according to Levenberg- Marquardt optiiation. Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 58

6 DOI:.18535/iecs/v6i.17 trainl(net,pd,tl,ai,q,ts,vv,tv) takes these inputs, net Neural network. Pd Delayed input vectors. Tl Layer target vectors. Ai Initial input delay conditions. Q Batch sie. TS Tie steps. VV Either epty atrix [] or structure of validation vectors. TV Either epty atrix [] or structure of validation vectors. and returns, net Trained network. TR Training record of various values over each epoch: Training occurs according to the trainl s training paraeters shown here with their default values [7]. 5. Algorith Using algoriths that were studied in the MATL AB progra in the field of artificial neural networks aft er training and application we are obtained for shapes where as illustrated in Figure () how they were arres ted on the body the recognied it and that the progra h as been training beforehand, either Figure () explains t he arrest and identification of body iage and the one who did not oents Hu unable to identify it. As for the fors that have not been previously trained by the progr a, it does not deal with the launching and gives the e ssage that it is not recognied body and that body is not a target as shown in Figure (1). Input Iige Figure () out progra result after treaining by Neuralnetwork R.Hue R.Neural network Input Iige Figure () Figure (1) out progra result after treaining by Neuralnetwork.5 R.Hue.4 R.Neural network Conclusions The ai of this work was to develop the software part of a Progra iaging syste that can autoatically identify obect in the field. This study resulted in developing a new recognition algorith which uses iage processing techniques based on artificial neural network ANN to identify shapes which we are training the progra for it. Tow recognition ethods, Hue oents, ANN ethods were used in this recognition syste. Note through a previous study that artificial neural networks were the best in ters of accuracy in the identification and discriination, where we have achieved a high percentage of known obects and the proportion of very few ratio virtually nonexistent copared to the ethod of oents Hu. As characteristic of these networks of high capacity for training and learning approach and previous values with subsequent References [1] GHassan.Satouf, Mahoud.Kaid," The Control Of Robot In Real Tie", Engineering Science Series, Res. J. of Aleppo,6//9. [] G.satoof, S. Jaklk," Use techniques of digital iages to deterine the coordinates of the Centre for Spatial body weight for robot vision syste", Engineering Science Series, Res. J. of Aleppo,8. [3] Y. Wu, et al," Siulation studies of data classification by artificial neural network", Potential applications in edical iaging and decision aking, Proc. SPIE Med. Iaging9,1998, [4] K. Prasad, D.C. Niga, A. Lakhotiya, D. Ure " Character Recognition Using Matlab s Neural Network Toolbox", B.I.T Durg, India, Vol. 6, No. 1, February, Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 59

7 DOI:.18535/iecs/v6i.17 [5] S. M. ASHAGHATHRA, Dr. Paul Weckler,"IDENTIFICATION OF PECAN WEEVILS THROUGH IMAGE PROCESSING", Oklahoa State University, May 8. [6]ROBI POLIKAR, Rowan University, Glassboro, New Jersey," PATTERN RECOGNITION", Wiley Encyclopedia of Bioedical Engineering, Copyright & 6 John Wiley & Sons, Inc. [7] Howard Deuth, Mark Beale," Neural Network Toolbox", User s Guide COPYRIGHT, by The MathWorks, Inc. [8] M.H Beale, M.T Hagan,H.B Deuth, " Neural Network Toolbox User's Guide", COPYRIGHT by The MathWorks, Inc. [9] David.S Touretk," Backpropagation Learning", /78: Artificial Neural Networks, Fall 6. Mark Hudson Beale Martin T. Hagan Howard B. Deuth Mahoud Mohaed Haoud Kaid, IJECS Volue 6 Issue Feb., 17 Page No.54-6 Page 6