Yangın ve Güvenlik Dergisi 200. Sayı (Temmuz-Ağustos 2018)

28 Yangın ve Güvenlik / Temmuz-Ağustos 2018 yanginguvenlik.com.tr KAPAK KONUSU / MAKALE 3. SINIRSEL AĞLARA DAYALI RETINA TANIMA SISTEMI Makalede ileri beslenen sinirsel ağ retina görüntüle- rinin tanınması için uygulanmıştır. Kullanılan NN (Neural Networks_sinirsel ağlar) giriş, gizli ve çıkış tabakaları içerir. Gizli tabaka ve çıkış tabakasının Nöronları (sinir hücreleri) içine Sigmois (S-biçimli) aktivasyon fonksiyonu kullanılmak- tadır. Gizli tabakanın sinir hücreleri hesaplandığında, daha sonra bunların aktivasyonu tüm aktivasyonlar sonunda çıkış tabakasına erişene kadar bir sonraki tabakaya beslenir. Her bir çıkış tabakası sinir hücresi özel bir sınıflandırma katego- risiyle birleştirilir. Çok tabakalı ileri beslemeli ağlarda Şekil 2, önceki tabakanın her bir sinir hücresi bir sonraki tabakaların sinir hücrelerine ağırlık katsayılarıyla bağlanır. Gizli ve çıkış tabakalarındaki her bir sinir hücresinin değerini hesaplama- da ağırlıkları alınan toplamların toplamı ve ön gerilmeleri (bayasları) ilk olarak alınmalı ve daha sonra sinir hücrelerinin aktivasyonu için aktivasyon fonksiyonu f (sum) (sigmoid_S biçimli fonksiyon) uygulanmalıdır. Anthemia hastalıklarının çıkartılan özellikleri nöral ağların girdileridir. Bu yapıda x1, x2, …, xm anthemia hastalıklarını karakterize eden girdi özel- likleridir. İki tabakalı sinirsel ağın “j” sayılı çıkışı (1) numaralı formülle belirlenir. Burada wij ağın girdi ve gizli tabakaları arasındaki ağırlık- lar, vjk gizli ve çıkış tabakaları arasındaki ağırlıklar, f nöron- larda kullanılan sigmois aktivasyon fonksiyonu, xi ise girdi sinyalidir. Burada k=1,..,n, j=1,..,h, i=1,..,m, m, h ve n ise giriş, gizli, ve çıkış tabakalarına karşılık gelen nöronların sayısıdır Sinirsel ağların aktivasyonundan sonra, sinirsel ağların parametrelerinin hazırlanması başlar. Nöral ağlar UCI kütüp- hanesinden alınan anthemia veri seti kullanılarak hazırlanır. Öğrenme safhasında sınıflandırmanın doğruluğunu değer- lendirmek için 10-katlı çapraz doğrulama kullanılır. Sinirsel ağların çıkışında gereken doğruluğu elde etmek için bir takım deneyler olmalıdır. Simülasyon gizli tabakadaki farklı sayıda nöronları kullanarak yapılır. Çıkış nöronlarının sayısı sınıfların sayısına eşit olan 8’di. Sinirsel ağların hazırlanması için geri yayılma algoritması uygulanmıştır 8. Sinirsel ağların hazırlanması olağan en küçük kareler fonksiyonunun mini- mize edilmesinden meydana gelmektedir (2). Burada O her sınıf için hazırlanma numuneleri sayısı, yd ve y girdi vektörü p’nin istenilen ve mevcut çıktılarıdır. Si- nirsel ağların hazırlanma parametreleri uygun bir sinirsel ağ modelini oluşturmak için yapılmaktadır. Sinirsel ağların (NN) ij jk w , v (i=1,...,m , j=1,...,h, k=1,…,n) parametreleri aşağıdaki formüller (3) kullanılarak ayarlanmaktadır. Burada, γ öğrenme oranı, i=1,...m; j=1,...h; k=1,…,n; m, h, n ağın giriş, gizli ve çıkış nöronlarının sayısıdır. Bunların türev- leri aşağıdaki gibi belirlenir (4): 4. SIMÜLASYON 4.1. Ön İşleme Retina tanıma sistemi tasarımı sinirsel ağlar dikkat alına- rak yapılmıştır. Başlangıç aşamasında retina görüntüleri dâ- hil görüntü veri tabanları tasarım maksadıyla kullanılmıştır. Bu nedenle makalede halka açık bir veri tabanı olan DRIVE (Digital Retinal Images for Vessel Extraction_Damar Özellik- lerin Çıkartılması için Dijital Retina Görüntüleri) veri tabanını Fahreddin Sadikoglu and Selin Uzelaltinbulat / Procedia Computer Science 102 (2016) 26 – 33 of retinal images is obtained. This image after normalization and enhancement is represented by the hat describes converted numeric values of the retinal image. For classification neural network is used. becomes the training data set for the neural network. The retina classification system includes two es: training mode and online mode. At first stage, the training of recognition system is carried out values of retina images. After training, in online mode, neural network performs classification nd patterns that belong to a certain retinal images. Fig. 1. A block diagram of the retina recognition system ork Based Retina R cognition System er feed-forward neural network is applied for identification of retina images. The used NN include nd output layers. The sigmoid activation function is used in the neurons of hidden and output layers. ns for the hidden layer are computed, their activations are then fed to the next layer until all the ally reach the output layer. Each output layer neuron is associated with a specific classification ultilayer feed-forward network at Fig. 2. each neuron of previous layers is connected the neurons of using weight coefficients. In computing the value of each neuron in the hidden and output layers one the sum of the weight d sums and the bias nd th n apply activation function f(sum) (the sigmoid lculate the neuron's ac ivation. The extracted features of the anthemia diseases are inputs of neural is structure, x 1 , x 2 , …, x m are input features that characterize the anthemia diseases. The j-th output of l networks is determined by the formula (1). 1 1 ( ( ) ) 1 ( ) 1 h m j k jk j ij i j i w h e y f v f w x f r e e ˜ ¦ ¦ ¦ ¦ (1) eights between the input and hidden layers of network, v jk are weights between the hidden and is the sigmoid activation function that is used in neurons, x i is input signal. Here k =1,.., n , j =1,.., h , nd n are the numbers of neurons in input, hidden and output layers, correspondingly. Fig. 2. Multilayer feed-forward network of neural network, the training of the parameters of neural network starts. Neural network is trained data set taken from UCI library. During learning the 10-fold cross validation is used for evaluation Pre-processing Retina image acquisition Feature vector Output Patterns Classification . : . : . : x 1 x 2 x m y 1 y 2 y k f ( ¦ ) f( ¦ ) (1) (2) (3) Şekil 2. Çok tabakalı ileri beslemeli ağ Fahreddin Sadikoglu and Selin Uzelaltinbulat / Procedia Computer Science 102 (2016) 26 – 33 etinal images is obtained. This image after normalization and enhancement is represented by the describes converted numeric values of the retinal image. For classification neural network is used. omes the training data set for the neural network. The retina classification system includes two training mode and online mode. At first stage, the training of recognition system is carried out lues of retina im ges. Aft r training, in online mode, neural network performs cla sification and erns that belong to a cert in retinal images. Fig. 1. A block diagram of the retina recognition system Based Retina Recognition System eed-forward neural network is applied for identification of retina images. The used NN include output layers. The sigmoid activation function is used in the neurons of hidden and output layers. for the hidden layer are computed, their activations are then fed to the next layer until all the reach the output layer. Each output layer neuron is associated with a specific classification ilayer feed-forward network at Fig. 2. each neuron of previous layers is connected the neurons of g weight coefficients. In computing the value of each neuron in the hidden and output layers one sum of the weighted sums and the bias and then apply activation function f(sum) (the sigmoid ate the neuron's activation. The extracted features of the anthemia diseases are inputs of neural ructure, x 1 , x 2 , …, x m are input features that characterize the anthemia diseases. The j-th output of tworks is determined by the formula (1). 1 1 ( ( ) ) 1 ( ) 1 h m j k jk j ij i j i w h e y f v f w x f r e e ˜ ¦ ¦ ¦ ¦ (1) hts between the input and hidden layers of network, v jk are weights between the hidden and he sigmoid activation function that is used in neurons, x i is input signal. Here k =1,.., n , j =1,.., h , n are the numbers of neurons in input, hidden and output layers, correspondingly. Fig. 2. Multilayer feed-forward network neural network, the training of the parameters of neural network starts. Neural network is trained a set taken from UCI library. During learning the 10-fold cross validation is used for evaluation Pre-processing Retina image quisition F ature vector Output Patterns Classification . : . : . : x 1 x 2 x m y 1 y 2 y k f ( ¦ ) f( ¦ ) (4) Fahreddin Sadikoglu and Selin Uzelaltinbulat / Procedia Computer Science 102 (2016) 26 – 33 of classification accuracy. There should be set of experiments in order to achieve required accuracy in the ne network output. The simulation is performed using different number of neurons in hidden layer. The numbe output neurons was 8 which was equal to the number of classes. The backpropagation algorithm is applied training of neural network 8 . Neural network training consists of minimizing the usual least-squares cost function 2 1 ) ( 2 1 y y E d O p ¦ (2) Where O is the number of training samples for each class, y d and y is the desired and current outputs of the p inp vector. The training of the neural network parameters has been carried out in order to generate a proper neural networks model. The parameters jk ij vw , (i=1,...,m , j=1,...,h, k=1,…,n) of NNs are adjusted using the following formulas (3). ( ) ( 1) ( ) ( ( 1) ( )); ( ) ( 1) ( ) ( ( 1) ( )); ij ij ij ij ij jk jk jk jk jk E t w t w t w t w t w E t v t v t v t v t v J O J O w w w w (3) where, J is the learning rate, i=1,...m; j=1,...h; k=1,…,n; m, h, n are the number of inputs, hidden and ou neurons of the network. The derivatives are determined as (4): ( ) ( ) ( ) y (1 ) ( ) ( ) ; ( ) ( ); y (1 ) v ; y (1 ) w ; d k k k k k j k jk k jk j k ij k j ij j d k k k k k jk j j ij k k j ij y E t E t y y y y v y v y y E t E t w y y w y y E t y y y y y y w w w w w w w w w w w w w w w w w w w w w ¦ ¦ (4) 4. Simulation 4.1. Pre-processing The design of retina identification system is considered using neural networks. At the start stage image data including retinal images is used for design purpose. For this reason in the paper we use DRIVE database whi publicly available database. The RGB retina images are transformed into grayscale images. A grayscale ima simply one in which the only colors are shades of grey. The reason for differentiating such image from any o sort of the color image is that less information needs to be provided for each pixel. In fact a gray color is on which the red, g ee , and blue components all have equal intensity in RGB space, and it is necessary to speci single intensity value for each pixel, as opposed to the three intensities needed to specify each pixel in all c images. Often, the grayscale intensity stored as an 8-bit integer giving 256 possible different shades of gray f black to white. Fig. 3. (a) shows us colored RGB retina image that is transformed to the grayscale image of r (b). In the paper for identification of retinal images we are using segmentation results of these images. In the r of segmentation the vascular represe tation of retinal images are ob ain d. (c) shows the results of segmentatio retinal image. The obtained image is scaled Fig. 4. and used for recognition purpose. Scaling is defined as increase or reduction of image size by a fixed ratio. We first smooth the image by convolution with a spat resolution. However, in a scale down by a specific factor in the respective directions, the image width to height r Fahreddin Sadikoglu and Selin Uzelaltinbulat / Procedia Computer Science 102 (2016) 26 – 33 of classification accuracy. There should be set of experiments in order to achieve required accurac netw rk output. The simulation is performed sing different number of neurons in hidden layer. output neurons was 8 which was equal to the number of classes. The backpropagation algorithm training of neural network 8 . Neural network training consists of minimizing the usual least-squares co 2 1 ) ( 2 1 y y E d O p ¦ (2) Where O is the number of training samples for each class, y d and y is the desired and current outputs o vector. T e tra ning of th neural etwork parameters has been carried out in rder to generate a prope networks model. The parameters jk ij vw , (i=1,...,m , j=1,...,h, k=1,…,n) of NNs are adjusted using the formulas (3). ( ) ( 1) ( ) ( ( 1) ( )); ( ) ( 1) ( ) ( ( 1) ( )); ij ij ij ij ij jk jk jk jk jk E t w t w t w t w t w E t v t v t v t v t v J O J O w w w w (3) where, J is the learning r te, i=1,...m; j=1,...h; k=1,…,n; m, h, n are the u ber of inputs, hidd neurons of the network. The derivati es are determined as (4): ( ) ( ) ( ) y (1 ) ( ) ( ) ; ( ) ( ); y (1 ) v ; y (1 ) w ; d k k k k k j k jk k jk j k ij k j ij j d k k k k k jk j j ij k k j ij y E t E t y y y y v y v y y E t E t w y y w y y E t y y y y y y w w w w w w w w w w w w w w w w w w w w ¦ ¦ 4. Simulation 4.1. Pre-processing The design of retina identification system is considered usi g neural networks. At the start stage i including retinal mages is used for esign pur ose. For thi reason in the pap r we use DRIVE dat publicly available database. The RGB retina images are transformed into grayscale images. A gray simply one in which the only colors are shades of grey. The reason for differentiating such image f sort of the color image is tha less inf rm tion needs to be provided for ea h pixel. I fact a gray which the ed, green, and blue components all have equal inte sity in RGB spac , and it is necessa single intensity value for each pixel, as opposed to the three intensities needed to specify each pi images. Often, the grayscale intensity stored as an 8-bit integer giving 256 possible different shade black to white. Fig. 3. (a) shows us colored RGB retina image that is transformed to the grayscale i (b). In the paper for identificatio of retinal images we re sing segmentation re ults of these image of segmentation the vascular representation of retinal images are obtained. (c) shows the results of s retinal image. The obtained image is scaled Fig. 4. and used for recognition purpose. Scaling is increase or reduction of image size by a fixed ratio. We first s mooth the image by co nvolution resolution. However, in a scale down by a specific factor in the respective directions, the image width Fahreddin Sadikoglu and Selin Uzelaltinbulat / Procedia Computer Science 102 (2016) 26 – 33 of classificatio accuracy. There should be set of experiments in order to achieve required accuracy i network output. The simulation is perfor ed using different number of neurons in hidden layer. Th output neurons was 8 which was eq al to the number of classes. The backpropagation algorithm is training of neural network 8 . Neural network training consists of minimizing the usual least-squares cost 2 1 ) ( 2 1 y y E d O p ¦ (2) Where O is the number of training samples for each class, y d and y is the desired and current outputs of t vector. T e training of the neural n twork parameters has been carried out in order to generate a proper networks model. The p rameters jk ij vw , (i=1,...,m , j=1,...,h, k=1,…,n) of NNs are adjusted using the f formulas (3). ( ) ( 1) ( ) ( ( 1) ( )); ( ) ( 1) ( ) ( ( 1) ( )); ij ij ij ij ij jk jk jk jk jk E t w t w t w t w t w E t v t v t v t v t v J O J O w w w w (3) where, J is the learning rate, i=1,...m; j=1,...h; k=1,…,n; m, h, n are the number of inputs, hidde neurons of the network. The derivatives are determined as (4): ( ) ( ) ( ) y (1 ) ( ) ( ) ; ( ) ( ); y (1 ) v ; y (1 ) w ; d k k k k k j k jk k jk j k ij k j ij j d k k k k k jk j j ij k k j ij y E t E t y y y y v y v y y E t E t w y y w y y E t y y y y y y w w w w w w w w w w w w w w w w w w w w w ¦ ¦ 4. Simulation 4.1. Pre-processing The design of retina identification system is considered using neural networks. At the start st ge im including retinal images is used for design purpose. For this reason in the paper we use DRIVE datab publicly available database. The RGB retina images a e transformed into grayscal images. A graysc simply one in which the only colors are shades of grey. The reason for differentiating such image fro sort of the color image is that less i formation eds o be prov ded for each ixel. In fact a gray co which the red, green, and blue components all have equal i tensity in RGB space, and it is necessary single intensity value f r each pixel, as opposed t he hr intensiti s ne ded t p cify each pixel images. Often, the grayscale intensity stored as an 8-bit integer giving 256 possible different shades black t white. Fig. 3. (a) shows us colored RGB eti a image that is transformed to the rayscale im (b). In the paper for identification of retinal images we are using segme ntation results of thes e images. of segmentation the vascular representation of retinal images are obtained. (c) shows the results of seg retinal image. The obtained image is scaled Fig. 4. and used for recognition purpose. Scaling is de