Dinamik algılayıcı öğrenme algoritması ile kenar saptamanın öğrenilmesi
Learning of edge detection using recurrent perceptron learning algorithm
- Tez No: 46271
- Danışmanlar: DOÇ.DR. CÜNEYT GÜZELİŞ
- Tez Türü: Yüksek Lisans
- Konular: Elektrik ve Elektronik Mühendisliği, Electrical and Electronics Engineering
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 1995
- Dil: Türkçe
- Üniversite: İstanbul Teknik Üniversitesi
- Enstitü: Fen Bilimleri Enstitüsü
- Ana Bilim Dalı: Belirtilmemiş.
- Bilim Dalı: Belirtilmemiş.
- Sayfa Sayısı: 58
Özet
ÖZET Bu tezde, Hücresel Yapay Sinir Ağlan için geliştirilen bir eğiticili öğrenme kuralı olan Dinamik Algılayıcı Öğrenme Kuralı (Recurrent Perceptron Learning Algorithm- RPLA ) ile genel olarak görüntü işlemenin özel olarak da kenar saptamanın öğrenilebileceği gösterilmiştir. Bu amaçla RPLA kullanan HYSA'lar için şablon öğrenme yazılımı CNNSLATW (Cellular Neural Network Supervised Learning Algorithm Tool for Windows) geliştirilerek bir kişisel bilgisayar ortamında benzetim düzeni elde edilmiştir. Bellek sorununu ortadan kaldırmak için CNNSLATW programı Visual C++ dilinde yazılmış ve böylece Windows ortamında çalışılarak yüksek çözünürlüklü görüntülerin işlenebilmesi sağlanmıştır. Yazılım RPLA ile görüntü işlemeye yönelik genel bir yazılımdır ve çeşitli uygulamalar (köşe saptama, kenar saptama, boşluk doldurma v.b.) için kullamlabilir. Ancak pek çok görüntü işleme uygulamasının ilk aşamasını teşkil ettiğinden bu çalışmada kenar saptama üzerinde durulmuştur. RPLA ile kenar saptama konusunda daha önce yapılan çalışmalarda küçük boyutlu, siyah-beyaz sentetik görüntüler kullanılmıştır. RPLA ilk kez bu çalışmada 256 x 256 görüntü birimi (pixel) boyutunda, gri seviyeli, gerçel görüntülerde kenar saptama için kullanılmıştır. Bulunan sonuçlar ile görsel örüntü tanıma ve anlamaya yönelik ara bir gösterilim elde edilmesi sağlanmıştır.
Özet (Çeviri)
SUMMARY LEARNING OF EDGE DETECTION USING RECURRENT PERCEPTRON LEARNING ALGORITHM Analog circuits have played a very important role in the development of modern electronic technology. Even in our digital computer area, analog circuits still dominate such fields as communications, power, automatic control, audio and video electronics because of their 'real-time' signal processing capabilities. Conventional digital computation methods have run into a serious speed batüeneck due to their serial nature. To overcome this problem, a new computation model, called 'artificial neural networks' has been proposed, which is based on some aspects of neurobiology and adapted to integrated circuits. The key features of artificial neural networks (ANN) are parallel and real-time processing. Although ANN becomes an important research area in the recent 10 years, the first works in this subject have been started at the beginnig of 1940's by McCullach, Pits, Hebb and Rosenblatt [1]. McCullach abd Pits showed it is possible to implement some functions like AND and OR using ANN. Then it gathers the interests of many researchers from various disciplines as well as from industrial intitutions, and almost all branches of engineering. But, since in these days silicon technology was not good anough to implement circuits with complexity needed for ANNs and digital computers known as an alternative for ANNs were developed to be operated at high speeds and their capacity and reliability are higher, from mid 60s to the beginning of 80s, related design and investment on ANNs were decreased. The advanced in silicon technology along with the fact that digital computers which have imcomparably high processing speeds in arithmetic operations could not show the same performance on such areas image processing, pattern recognation where the problems with incomplete data and/or not well defined have attracted a great deal of scientists from various disciplines to the field again. In the past three decades a number of neural network architectures have been developed. The architectures have been inspired both by the principles governing VIbiological neural systems and well-established theories of engineering and fundamental sciences. The best known neural network models are Grossberg's Adaptive Resonance Theory: ART, Widrow*s ADAptive LINEar Element: ADALİNE, Hopfield Network, Multilayer Perceptron, and Kohonen's Self Organizing Feature Map. Most of the widely applied neural networks fall into two main classes: 1) memoryless neural networks and 2) dynamical neural networks. From a circuit theoretical point of view, the memoryless neural networks are non-linear resistive circuits, while the dynamical neural networks are non-linear R-L-C circuits. A memoryless neural network defines a non-linear transformation from the space of input signals into the space of output signals. Such networks have been successfully used in pattern recognition and several problems which can be defined as a non-linear transformation between two spaces. As İn the Hopfield network and Cellular Neural Network, dynamical neural networks have usually been designed as dynamical systems where the inputs are set of some constant values and each trajectory approaches one of the stable equilibrium points depending upon the initial state. Some useful application of these networks includes image processing, pattern recognition and optimization. From the learning point of view, neural network models can also be classified into three main classes: ' 1) Supervised Learning 2) Unsupervised Learning 3) Non-Learning One of the most important contemporary interest is developing neural network models which not only provide new features in information processing, but also are more appropriate to hardware realizations. The Cellular Neural Network (CNN) model propased by Chua and Yang in 1988 is one of the remarkable developments in neural network theory and is a significiont contribution to the development of a variaty of image processing applications. The basic circuit unit of Cellular Neural Network is called a cell. It contains linear and nonlinear circuit element, which typically are linear capacitors, linear resistors, linear and nonlinear controlled sources, and independent sources. Any cell in a CNN is connected only to its neighbor cells. The neighborhood defined by following metric: A A d(i,j;i, j) = max A i- i A j-j Where (i,j) is the vector of integers indexing the cell C(ij) in the i th row j it column of the 2-dimensional array. The system of equations describing a CNN with the neighborhood size of one is given in (l)-(2). VllXU=“A'XU + E Wm.nyi+mj+n + 2>m,”Uj+m>j+n + I (1) m.n e -1,0,1 ' m,n6 -1,0,1 yu =f(x.,j) = rkj + il-Ki“1! (2) Where, A, I, wpJ and Zp, e R are constants. Theoretically, the network structure of CNN can be defined of any dimension, but in this thesis focuses the attention on one of the image processing problems, namely edge detection, so CNN is defined to be a single-layer, first order network made of regularly spaced cells inter connected so as to form a two dimensional grid. A CNN is completely stable if the feedback connection weights wp, are symmetric. Throughout the thesis, the input connection weights z^, are chosen symmetric for reducing computational costs while the feedback connection weights wpi are chosen symmetric for ensuring the complete stability, i.e., def def def w^.j = wu = a!, w_1)0 = w1>0 = %, w_u = w1(_, = a3, def def wo,-ı= w0jı =a4, w0>0 = a5; def def def def def Z-l,-l = ^.l = ”1 ' Z-l,0 = Zl,0 =“2 » Z-l,l = ^.-l = ”3 ' Z0,-l = Z0,l = D4 ' ^,0 = "5. Digital image processing systems are designed to recover useful information from one or more images of a scene. To facilitate the analysis of image, image processing systems are considered to comprise three processing levels, which are commonly referred to as low level, indermadiate level, and high level. Edge detection is one of the most important parts of low level processing. This is obvious from human vision for which line drawing and cartoons provide sufficient information for object identification. By detecting edges, an image processing system (biological or digital) finds occluding countours of objects. The edge detection process serves simplify the analysis of images by drastically reducing amount of data to be processed, while at the same time serving useful structural information about object boundaries. V1UIn digital image processing basically two methods are used to detect edges in gray scale images. The first approximates the gradient and uses its length to determine edges. The second method, known as parallel edge detection and used in high speed digital image processing consists of three steps: i) Correlation of the image with a local difference operator, yielding a measure for the strength of a gray scale discontinuity. ii) Thresholding of the correlated image in order to classify edges. iii) Elimination of noise or bridging of gaps. A difficulty of this procedure is to determine on appropriate operator and a corresponding threshold in order to get a relevant image. This was often done heuristically, or in some cases following an analytical approach proposed by Canny. As an alternative, neural networks have been suggested. By using neural networks, appropriate training images can be designed and the operator and threshold can be learned. The foregoing studies show that CNN is capable of earring out some complicated information processing tasks, and is specifically very successful in relatively low-level image processing applications such as edge detection, corner detection, hole filling and noise filtering. The main concern in CNN applications is to determine the inter connection weights which perform the desired image processing. Various methods exist as solutions of this problem. Learning algorithms are an important part of these solutions. Learning is performed through the modification of the interconnection weights according to an optimality criterion, depending on the samples chosen from the training set which is constituted of some pre-determined input signals and the corresponding desired outputs (if known) as in the supervised learning. The supervised learning of the steady-state outputs in completely stable CNNs is a constrained optimization problem, where the objective function represents the error between the actual and desired outputs. The constraints are originated from the basic operation principles of CNN and are due to the desired design requirements such as the bipolarity of the steady-state outputs and the complete stability. Recently, a supervised learning algorithm has been proposed for the completely stable CNNs. It resembles the Perceptron learning algorithm, and hence is called the Recurrent Perceptron Learning Algorithm (RPLA), since applied to a dynamical neural network, CNN. RPLA is developed for finding the interconnection weights and threshold of CNN to realize an input- steady-state output map described by a set of training samples. In this thesis, RPLA is used for finding the optimal template coefficients and the threshold for the edge detection. To achieve this, the input-desired output pairs in the training set are applied to the CNN in an order. The images to be processed input to the IXnetwork from the external inputs and initial states. Desired output is^hosen as the image which represents the edges of the image to be processed. After each iteration, weight vector w is updated according to the error which shows the difference between desired and actual outputs. In this thesis, it is showed that edge detection can be learned using Recurrent Perceptron Learning Algorithm (RPLA) which is recently proposed for supervised learning of the steady-state outputs in completely stable Cellular Neural Networks. To achieve this work, a template learning software CNNSLATW (Celluar Neural Network Supervised Learning Algorithm Tool for Windows) is developed and the images which are gray scale, 256x256 pixel size are processed for the first time in the literature. To overcome memory problem CNNSLATW was written in Visual C++. After giving a general overview of the ANN concept in Chapter 2, the CNN are presented in Chapter 3 along with a brief discussion of its up-to-date applications to image processing. Chapter 4 includes a brief information about learning in ANNs and the Recurrent Perceptron Algorithm (RPLA) proposed recently which is used for edge detection in this thesis. In chapter 5, after giving information about some conventional edge detection algorithms, the learning of edge detection using RPLA is presented. This chapter also includes the simulation results of CNNSLATW program.
Benzer Tezler
- Image restoration using dynamical neural networks
Dinamik yapay sinir ağları ile görüntü onarımı
ATA ÜNAL
Yüksek Lisans
İngilizce
1994
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiDOÇ.DR. CÜNEYT GÜZELİŞ
- Hücresel yapay sinir ağları için iki öğrenme algoritması ve görüntü işleme uygulamaları
Two learning algorithms for cellular neural networks and their image processing applications
SİNAN KARAMAHMUT
Yüksek Lisans
Türkçe
1994
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiDOÇ.DR. CÜNEYT GÜZELİŞ
- Design of algebraic and dynamical PWL and PWC neural networks for classification
İşaret sınıflandırma ve işaret işleme amaçlı cebrik ve dinamik PPS ve PPD modeller
İBRAHİM GENÇ
Doktora
İngilizce
2007
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiElektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
PROF. DR. CÜNEYT GÜZELİŞ
- Kablosuz algılayıcı ağlar ile bal arısı kolonilerinin izlenerek arıcılık potansiyelinin tahmini için zeki karar destek sistemi geliştirilmesi
Development of the intelligent decision support system for the estimation of beekeeping potential by monitoring honey bee colonies with wireless sensor networks
AHMET ALBAYRAK
Doktora
Türkçe
2018
Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolKarabük ÜniversitesiBilgisayar Mühendisliği Ana Bilim Dalı
PROF. DR. RAİF BAYIR
- A hybrid deep learning metaheuristic model for diagnosis of diabetic retinopathy
Diyabetik retinopatinin tanısı için hibrit bir derin öğrenme meta-sezgisel modeli
ÖMER FARUK GÜRCAN
Doktora
İngilizce
2022
Endüstri ve Endüstri Mühendisliğiİstanbul Teknik ÜniversitesiEndüstri Mühendisliği Ana Bilim Dalı
DR. ÖĞR. ÜYESİ ÖMER FARUK BEYCA