Genetic Optimization of Neural Networks for Person Recognition Based on the Iris

This paper describes the application of modular neural network architectures for person recognition using the human iris image as a biometric measure. The iris database was obtained from the Institute of Automation of the Academy of Sciences China (CASIA). We show simulation results with the modular neural network approach, its optimization using genetic algorithms


Introduction
The recognition of persons using biometrics is a problem that has been considered by many researchers [1][2][3][4].Biometrics plays an important role in public safety and to accurately identify each individual to distinguish them from each other [5].This problem has been studied more thoroughly in recent years thanks to advances in computational power that have allowed the implementation of more complex algorithms using different techniques [6], [7].Biometric identification systems are those based on physical characteristics or morphology of human beings to perform some kind of recognition [8], [9].
Traditional systems used in accessing control are based on magnetic cards, card systems with bar code systems capture key or a combination.These systems involve the use of a card that must be carried always and which is not exempt from being lost, damaged, be stolen or forged, thus security is more vulnerable to failure.For this reason, systems that are more robust and with higher reliability are needed to avoid the problems mentioned above.Pattern recognition systems based on neural networks have been given recently considerable interest [10][11][12][13][14][15].
There are different techniques and methods that can be used for feature extraction, and today it is easier to recognize a person by the existing biometric methods.For example, a person can be recognized for its iris, fingerprints, and face, recognizable by his voice, signature, hand geometry, ear, vein structure, retina, facial thermography and others that exist [16][17][18][19].At this moment, biometric methods have been implemented using different devices to create patterns and generate the code that identifies the persons [20][21][22][23].
Biometrics refers to an identification and authentication technology that is transforming a biological, morphological, or behavioral characteristic into a numerical value.Its aim is to attest to the uniqueness of a person from far irrepressible immutable part of the body [5].Another definition mentions that biometrics is based on the premise that each individual is unique and has distinctive physical traits or behaviors, which can be used to identify or validate [24].
Within the large field of biometrics where one can highlight, fingerprint recognition, retinal and voice, among others, we can highlight the iris recognition as a biometric tool for person recognition in a unique and highly accurate fashion [1], [2].
This paper presents research work on integrating results of a modular neural network using the CASIA database, and obtaining the best identification when using type-1 fuzzy logic integrators developed by the genetic algorithms.Optimization of the neural networks was performed with genetic algorithms (GAs), which are essentially a method that creates a population of individuals to find the most appropriate one by simulating evolution [25][26][27][28].This process is based on natural selection by using operators such as the crossover and mutation to create new individuals.The modular neural network architectures and the chromosomes produced by the genetic algorithms with the best parameters found for the network were tested for their performance and operation, and the results of the different integrators, such as the gating network and type-1 fuzzy logic integrators, were compared for this problem.

Research Method
Neural networks are composed of many elements (Artificial Neurons), grouped into layers that are highly interconnected (with the synapses), which are trained to react (or give values) in a way you want to input stimuli.These systems emulate in some way, the human brain.Neural networks are required to learn to behave (Learning) and someone should be responsible for the teaching or training (Training), based on prior knowledge of the environment problem [7], [5].
A neural network is a system of parallel processors connected together as a directed graph.Schematically, each processing element (neuron) of the network is represented as a node.These connections provide a hierarchical structure trying to emulate the physiology of the brain for processing new models to solve specific problems in the real world.What is important in developing neural networks is their useful behavior by learning to recognize and apply relationships between objects and patterns of objects specific to the real world.In this respect neural networks are tools that can be used to solve difficult problems [29], [8], [30].Artificial neural networks are inspired by the architecture of the biological nervous system, which consists of a large number of relatively simple neurons that work in parallel to facilitate rapid decision-making [24].
Fuzzy logic was proposed for the first time in the mid-sixties at the University of California Berkeley by the brilliant engineer Lotfi A. Zadeh [31], [32].Who proposed what it's called the principle of incompatibility: "As the complexity of system increases, our ability to give precise instructions and build on their behavior decreases to a threshold beyond which the accuracy and meaning are mutually exclusive characteristics."Then introduced the concept of a fuzzy set, under which lies the idea that the elements on which to build human thinking are not numbers but linguistic labels.Fuzzy logic can represent the common knowledge as a kind of language that is mostly qualitative and not necessarily a quantity in a mathematical language by means of fuzzy set theory and the characteristic functions associated with them [32].
Fuzzy logic has gained a great reputation for the variety of applications, ranging from control of complex industrial processes to the design of artificial devices for automatic deduction, through the construction of household electronic appliances and entertainment as well as diagnostic systems [33][34][35][36][37][38].
Fuzzy logic is an area of soft computing, which allows one computer system to the reason for the uncertainty [31].This corresponds, in the real world, to many situations where it is difficult to decide unequivocally whether or not something belongs to a specific class [39][40][41][42].Fuzzy logic is a useful tool for modeling complex systems [43][44][45][46][47][48].However, it is often difficult for human experts to define the fuzzy sets and fuzzy rules used by these systems [36].This is particularly true for type-2 fuzzy systems that use uncertain membership functions and that have recently been applied to many real-world problems [49][50][51][52][53][54][55][56][57].
Genetic algorithms were introduced by the first time by a professor of the University of Michigan named John Holland [31], [5].A genetic algorithm, it is a mathematical highly parallel TELKOMNIKA Genetic Optimization of Neural Network for Person Recognition Based on algorithm that transforms a set of mathematical individual objects with regard to the time using operations based on evolution.The Darwinian laws of reproduct be used, and after having appeared of natural form a series of genetic operations between (among) individuals that stand out for the sexual recombination [ one generation to another a serie operators are selection, crossover being a chain of characters (letters or numbers) of fixed length that adjust to the model of the chains of chromosomes, and one associates to them with a certain mathematical function that reflects the fitness.
There exists a diversity of voting method, fuzzy integration, and gating networks [ section (for illustrative purposes) on the gating network method.
Integration by Gating Network tasks learned through the modules Gating Network are: best overall performance classifiers, need not be the same type There are several implementations of the important is by nature of using neuron to evaluate the performance gating network is based on a networks of experts [5].In Figure

Iris Image Preprocessing
Due to the unique, stable identification based on the iris The idea of using iris patterns ophthalmologist Frank Burch.Safir, American ophthalmologists, patented the system, led to contact with John he developed the necessary algorithms [7].These algorithms, patented by of all iris recognition systems Various studies carried out for which uses neural networks and identifier is perhaps one of the most appearance of the iris.This identifier The database of human China [58].This institution has several the database, which consists of people the total database.The algorithm that transforms a set of mathematical individual objects with regard to the time using operations based on evolution.The Darwinian laws of reproduction and survival of the fittest can be used, and after having appeared of natural form a series of genetic operations between (among) individuals that stand out for the sexual recombination [25], [26].For the a series of genetic operators are applied.The most commonly used selection, crossover and mutation [15].Each of the individuals is in the habit of being a chain of characters (letters or numbers) of fixed length that adjust to the model of the chains of chromosomes, and one associates to them with a certain mathematical function that here exists a diversity of methods of integration or aggregation of voting method, fuzzy integration, and gating networks [25].However, we concentrate in this section (for illustrative purposes) on the gating network method.
Gating Network: in this case decomposition of a learning task the modules of cooperation is performed.The benefits best overall performance, reuse of existing patterns heterogeneity same type; different features can be used for different implementations of the modular neural network of using the gating network.In some cases, this corresponds performance of the other modules of experts.Other embodiment of the on a neural network trained with a different data set of the most foreign to people, as among us do not This identifier is one of the most accurate among biometric systems human iris is from the Automation Institute of the Academy has several databases of iris, and we used in this work which consists of 14 images per person (7 of each eye), we used The image dimensions are 320x280 pixels, the format is for training and 6 for testing.
.… (Patricia Melin) 311 algorithm that transforms a set of mathematical individual objects with regard to the time using ion and survival of the fittest can be used, and after having appeared of natural form a series of genetic operations between For the passage from most commonly used Each of the individuals is in the habit of being a chain of characters (letters or numbers) of fixed length that adjust to the model of the chains of chromosomes, and one associates to them with a certain mathematical function that of information, like ].However, we concentrate in this learning task into sub benefits of working with heterogeneity expert different classifiers.neural network, but the most corresponds to a single Other embodiment of the data set for training the network integrator is presented.
in 1936 by the Leonard Flom and Aran His interest in developing the at the University of Harvard so he pattern of the iris published in [14], are the basis Ahmad Sarhan [7], based identification.The iris as an us do not recognize the biometric systems [7]. of the Academy Sciences of and we used in this work version 3 of ), we used only the first 77 , the format is JPEG, and In the pre-processing stage, noise removal, in order to extract filters on it, this in order to help images (see Figure 2).

Statement of the Problem and Proposed Method
We studied several methods networks for person recognition methods for response integration of the winner takes all.Figure 5 shows

Statement of the Problem and Proposed Method
methods of fuzzy integration that can be applied recognition using biometric iris images as well to develop for response integration of the modular neural network, such as the gating 5 shows the general architecture that was used in this work and this allows you to solve problems matrix to represent neural networks binary vector) to a single output value The first module is specialized 33 for the second module and network, and training is conducted parameters of the (MNN) are shown on Table 1 In this case each module for each of the 3 modules, a study are appropriate for the ANN training used.This will depend on learn faster or slower and therefore adapt to the learning method second and third, module, unlike each module are powered by uniform.The main advantage which can be helpful because on the outcome of the integration, which takes results.The network consists of step of the network, and once connections of the layers of network with the gating network (the result of the integration of Table 1 M. Training Trainscg: Scaled conjugate Traingda: Gradient descent with momentum and adaptive learing Traingdx: Gradient descent with adaptive learning factor Table 2. Results of the Table 2 shows results obtained from modules, the best methods that obtained a trainscg, which achieved a recognition with a time of 3minutes and 30 seconds 8000, 5000, and 8000 epochs trainings were performed without results were not satisfactory, because there are 2, June 2012 : 309 -320 solve problems that are not linearly separable.The perceptron networks and it is a tertiary discriminator that traces output value f (x) (a single binary value) through the matrix specialized for the first 33 people who were vectorized and the last 33 for the third module, forming conducted according to the sequence of modules 1 st are shown on Table 1.module is fed with the same information, to find a suitable a study in an empirical way was done to know how many ANN to have a learning that is acceptable according will depend on the number of neurons used for such training therefore have a good or bad learning, for this reason is along with the epochs.To know which structure is right unlike ensemble neural networks, modular neural networks by vectors of different data, leading to architectures that advantage of this method is that each module produces because the training is not leaved to one expert module.This integration, which takes into account that each module consists of three modules, the weight distribution is random , and once this is completed the network will adjust the weights the neural network.The results of each module network integrator and arbitrary parameters, without the integration of the 3 modules), are shown in Table 2.  as the neurons.For this reason rate of recognition, and thus obtain the appropriate hidden layers, and once we get many times the genetic algorithm module of the modular neural network with gating network integration and with pre (the result of the integration of the 3 modules) are shown in Table 3. Results of the Table 3 shows the results compression, which can be seen same parameters of the training traingda, traingdx, trainscg with 50 seconds, 3 minutes 36 seconds and epochs respectively, and a goal error of neurons.For this reason it was chosen to use a genetic algorithm to find obtain the appropriate type of training, the number of neurons and we get these parameters take the mean and standard deviation the genetic algorithm can find a high percentage of recognition.The results of each ural network with gating network integration and with pre (the result of the integration of the 3 modules) are shown in Table 3. 315 find an appropriate number of neurons and and standard deviation of how of recognition.The results of each ural network with gating network integration and with pre-processing preprocessing image processing and of recognition with the high percentage were with a time of 4minutes with 8000, 5000 and 5000 modular neural network

Optimization of the Architecture
To optimize the modular neural network a genetic algorithm was used to find the optimal architecture and an appropriate recognition rate.The general scheme of the MNN indicating the three modules and Once it was verified that the GA found a good optimization result, it was decided to run the GA 10 times to find the standard deviation and average of the results for the neurons, methods, and the number of layers.The summary of the training behavior for the GA was obtained with the results of the experiments and this forms the basis of possible comparisons with other optimization approaches.The parameters of the chromosome tha GA are shown in Table 4.
The real chromosome 250 neurons, which varied over a range training methods are shown on Table 5.

Results and Analysis
The results of executing 10 are shown in the following Tables: Table 8.The results for the optimization of shown in Table 6.

Table 6. Results of
The results for the optimization of shown in Table 7.The results for the algorithm are shown in Table performed with the genetic algorithm the average in each generation better error (B /E /GA) found by the 2, June 2012 : 309 -320

Optimization of the Architecture
To optimize the modular neural network a genetic algorithm was used to find the optimal architecture and an appropriate recognition rate.The general scheme of the MNN ndicating the three modules and the integrator is shown in Figure 6.
Once it was verified that the GA found a good optimization result, it was decided to run the GA 10 times to find the standard deviation and average of the results for the neurons, methods, and the number of layers.The summary of the training behavior for the GA was obtained with the results of the 10 experiments and this forms the basis of possible comparisons with other optimization approaches.The parameters of the chromosome that were used chromosome is composed of 3 layers {1, 2, 3}, and each layer is over a range from 0 to 250 values and 4 training methods are shown on Table 5.To optimize the modular neural network a genetic algorithm was used to find the optimal architecture and an appropriate recognition rate.The general scheme of the MNN Once it was verified that the GA found a good optimization result, it was decided to run the GA 10 times to find the standard deviation and average of the results for the neurons, methods, and the number of layers.The summary of the training behavior for the GA was obtained with the results of the 10 experiments and this forms the basis of possible comparisons t were used in the layer is composed of training methods.The  As comparison of results, we can mention that in [47] a 93.33% recognition rate was achieved, while in this work we were able to obtain recognition rates of 99.76%.This fact shows that the proposed approach can outperform similar neural approaches in the literature for iris recognition.

Conclusion
The best result for person recognition obtained through a set of modular with 116 and 117 neurons in 112 to 114 neurons in the third genetic algorithm run is as follows: for the first module, average error of 0.0152, for the 2nd the third module of 98.59% with an integration architecture the average membership functions, the validation of this fuzzy integrator with Gaussian membership average a recognition of 99.52 MFs the average recognition was average recognition was 99.76 satisfactory, it was decided functions of this response integrator of the modular neural network.chosen to optimize this integration system was a algorithm a better recognition rate was achieved integration system were obtained in the modular neural network.As comparison of results, we can mention that in [47] a 93.33% recognition rate was while in this work we were able to obtain recognition rates of 99.76%.This fact shows that the proposed approach can outperform similar neural approaches in the literature for iris for person recognition using the iris biometric measurement modular neural network architectures with 3 layers in each neurons in the first hidden layer, 116 and 113 in the 2nd the third hidden layer.The average percentage in each generation is as follows: for the first module, a recognition rate of the 2nd module of 97.98%, with an average error of % with an average error of 0.0141.For validation of the architecture the average was 98.48%, for the fuzzy integrator with , the validation of this integrator was on average 99.37 Gaussian membership functions, the validation of this integrator 99.52%.In the optimized cases, for the validation with MFs the average recognition was 99.64%, and for the validation with Gaussian 99.76%.Since initially the results with fuzzy integration were not , it was decided to apply an evolutionary approach to optimize ponse integrator of the modular neural network.The evolutionary method integration system was a genetic algorithm.After applying the recognition rate was achieved, because better results with were obtained in the modular neural network.
.… (Patricia Melin) As comparison of results, we can mention that in [47] a 93.33% recognition rate was while in this work we were able to obtain recognition rates of 99.76%.This fact shows that the proposed approach can outperform similar neural approaches in the literature for iris biometric measurement was layers in each module, hidden layer, and each generation of the a recognition rate of 98.48%, with an average error of 0.0202, and for or validation of the gating network with Triangular type 99.37%, and for the this integrator was on for the validation with triangular type Gaussian type MFs the integration were not optimize the membership The evolutionary method After applying the genetic results with the optimized

Figure 1 .
Figure 1.Representation of the gating network integration method 8 images were used for training ISSN: 1693-6930 Genetic Optimization of Neural Network for Person Recognition Based on .

Figure 1 a
scheme of the gating network integrator is presented Representation of the gating network integration methodIris Image Preprocessing, stable and accessible characteristics of iris patterns iris pattern has become one of the most reliable patterns to identify people was first proposed Burch.However, it was not until 1987, when Leonard ophthalmologists, patented the concept of Burch.His interest John G. Daugman, then a professor at the University of the necessary algorithms for biometric recognition through the pattern , patented by Daugman in 1994 and partly published in [ that exist today.carried out for iris recognition, as the work of M. Ahmad neural networks and the cosine transform for iris-based identification.

Figure 3
Figure 3 shows the resulting image, maximum and minimum.As shown outer parts of the circle, leaving them of the iris, and the other way the image for the network.

Figure 3 .
Figure 3. Result of applying different techniques of

4 .
Application of the wavelet transform in 2D persons we have 14 samples, which means that there are 7 and 7 samples of the left eye, of which we took 8 sample to validate that the images will be recognized according to of vectors, the first containing the training images 33), 264 vectors in the matrix of training and (6 * validation matrix.The same was done for persons (34-66) and (67 the modules of the modular neural network. General architecture of the modular neural network neural network consists of 3 modules, each module is an artificial neural network (ANN) consisting .…(Patricia Melin) 313 there are 7 pictures or sample images for according to the 8 above, and the other the (6 * 33), 198 images 66) and (67-99), for each of to modular neural develop alternative network, such as the gating network and this work.neural network , each module consists of a of multiple layers, network without preprocessing in the results obtained from various trainings performed in that obtained a high recognition percentage were traingda a recognition above 90%, with 90.18, 90.91 and 90.40 3minutes and 30 seconds, 3 minutes 15 seconds and 3 minutes epochs respectively, and a target error of 0.00001 for the without pre-processing and with an uncompressed , because there are parameters that have very large

Figure 6 .
Figure 6.General of the Modular Neural Network in the image with preprocessing shows the results of the trainings conducted, pre-image can be seen that the rates increased by almost 4% of recognition with of the training carried out previously, the methods with high percentage with 95.89, 94.97 and 93.33% respectively with a time of 36 seconds and 4 minutes 22 seconds, with 8000 and a goal error of the ANN of 0.00001.General architecture of optimized modular neural network .…(Patricia Melin) executing 10 times the genetic algorithm (GA) for each of the modules are shown in the following Tables: Module 1 in Table 6, Module 2 in Table 7 optimization of Module 1 for 10 runs of the genetic algorithm are of the genetic algorithm run by generations for module optimization of Module 2 for 10 runs of the genetic algorithm are The results for the optimization of Module 3 for 10 runs of the algorithm are shown in Table 8.In Tables 6, 7 and 8 we show the results genetic algorithm, with 20 generations, 10 individuals and each generation run and the standard deviation of each generation found by the genetic algorithm (GA), better training method ISSN: 1693-6930 Gradient descent with momentum and adaptive learning factor GA) for each of the modules 7 and Module 3 in genetic algorithm are for module 1 genetic algorithm are runs of the genetic the results of the trainings 10 runs, showing each generation and run, training method and TELKOMNIKA Genetic Optimization of Neural Network for Person Recognition Based on execution time of the 10 runs.architectures of the different training ISSN: 1693-6930 Genetic Optimization of Neural Network for Person Recognition Based on .
10 runs.Once the 10 runs of the GA were achieved, we obtained training (150 per run). of the genetic algorithm run by generations for module Results of genetic algorithm run by generations for module 3

Table 4 .
Parameters of the chromosome for the GA

Table 5 .
Training for the GA

Table 7 .
Results of