12759

11675

كارشناسي ارشد

مخابرات سيستم

اصفهان: دانشگاه صنعتي اصفهان، دانشكده برق و كامپيوتر

1396

دوازده، 66ص.: مصور، جدول، نمودار

ص.ع. به فارسي و انگليسي

بهزاد نظري، سعيد صدري

علي يوسفي

1396/06/15

كتابنامه

برق و كامپيوتر

مهندسي برق و كامپيوتر

ID11675

Improvement of spiking neural networks for image segmentation and data classification Mozhgan Mokhtari mozhgan mokhtari@ec iut ac ir Date of Submission 2017 05 29 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Dr Behzad Nazari Dr Saeed SadriAbstractNeuronal signal consists of short voltage pulses called action potentials or spikes The sequence ofaction potentials contains the information that is conveyed from one neuron to others In biologicalnervous systems the transmitted information is usually encoded in the frequency of spiking and orin the timing of the spikes A spiking neuron transmits information by the timing of the spikes Spiking neural networks are networks of spiking neurons that their inputs and outputs are spikefiring times At first the pulse neural networks were introduced and modeled on a single layer inwhich neurons did not need to be trained In these models neurons are sensitive to intensity levelof pixles of an image as inputs and produce pulses based on the amount of stimulation at differenttimes These networks were used in image processing effectively In this thesis as an applicationwe make use of simultaneous feature of activating neurons with the same stimulation intensity Thepulse neural network is used for binarization of image of documents The purpose of this study isto compare pulse neural network performance with other document binarization techniques In last few years by providing appropriate training methods spiking neurons are also trained Itis shown developing an effective learning method is a challenge in the study of spiking neural net works One of the most widely used learning methods in this field is gradient descent based learningmethod However in these methods due to complexity of computations related to the hidden layer the network requires a lot of learning epochs In this thesis by modification of the network errorfunction and according to the input coding the number of required layers of the network is reducedto two Accordingly by removing the hidden layer the convergence speed is improved Keywords 1 Spiking neural networks 2 Multi spike learning 3 Single spike learning 4 Document image binarization

بهزاد نظري، سعيد صدري

علي يوسفي