Hakan Tora

Emotion classification using hidden layer outputs

Neural network (NN) with Multi-Layer Perceptron (MLP) is a supervised learning algorithm composed of artificial neurons. Multilayer NN is capable of solving nonlinear classification problems such as emotion identification by using facial expressions that is presented in this paper. Hidden layer outputs of NN provide useful information about facial appearance. This study addresses that without fully training NN hidden layer outputs can be used as feature. It is shown that an acceptable recognition rate is obtained by means of hidden layer outputs.

An alternative method for cell counting

Cell counts and classification of the cells play an important role in the field of microbiology and cell biology. Although there exists many counting processes for cells of interest in suspension, the most basic cell counting process is performed by a person via the microscope. For counting cells the simplest, widely used and the most economic method is the use of hemocytometer counting. In this study, the hemocytometer counting was used but the the cells were counted by a proposed image based approach. The developed technique herein uses neural network along with the Hough transform.

Higher order statistical analysis of Turkish phones

In this study, histograms of Turkish phones were examined using higher order cumulants. As is known, phones constituting a language, are composed of letters as vowels and consonants. These letters can also be grouped as voiced and unvoiced phones. It is observed that unvoiced letters show a Gaussian-like distribution and result in small values of skewness and kurtosis. On the other hand, vowels and voiced consonants lead to a non-Gaussian distribution. Voiced and unvoiced phones are related with their skewness and kurtosis values. It is empirically shown that higher order cumulants are likely to be a feature in describing Turkish phones.

Performance evaluation of self organizing neural networks for clustering in ESM systems

Electronic Support Measures (ESM) system is an important function of electronic warfare which provides the real time projection of radar activities. Such systems may encounter with very high density pulse sequences and it is the main task of an ESM system to deinterleave these mixed pulse trains with high accuracy and minimum computation time. These systems heavily depend on time of arrival analysis and need efficient clustering algorithms to assist deinterleaving process in modern evolving environments. On the other hand, self organizing neural networks stand very promising for this type of radar pulse clustering. In this study, performances of self organizing neural networks that meet such clustering criteria are evaluated in detail and the results are presented.

Real time infrared image enhancement

This study evaluates the implementation of Balanced Contrast Limited Adaptive Histogram Equalization (BCLAHE) for infrared images (IR) on an embedded platform. The aim was to achieve real time performance for the operator display target application. The system configured for this aim is a dual processor media application device OMAP3530, which consists of an ARM and a DSP processor. System is configured so that hardware sources are used efficiently and various performance improvement techniques are investigated. Performance analysis is done over IR images with different dynamic range.

Tree based neural network design for emotion identification analysis

Emotion identification analysis became popular research area nowadays. It can be used in many areas such as physiology, education, murder squad, tendency to crime to get a clue about mental signals of a person. Facial expressions are kind of communication channels that carry sense signals. Therefore, they are as important as speech and body movement. Sometimes they are much more meaningful because of their naturalness. That is why it is appreciated to work on automatically recognition of facial expressions. This paper proposes an approach to recognize facial expressions by using neural network. Using one unit neural network is enough to recognize the facial expressions but using a tree structure neural network increases the accuracy of the results and the performance of the testing set. In this study, it is proposed a tree network architecture which yields better recognition performance.

Naturalness analysis of the speech synthesized by a TTS card

It is known that the performance of a developed text-to-speech (TTS) synthesis system is assessed by subjective tests. These assessments are usually based on the intelligibility and naturalness of the synthesized speech. In this study, an investigation on relating these subjective test results, thus the naturalness of the synthesized speech, to which acoustic features is accomplished. Consequently the features which will increase the performance while synthesizing the speech are determined. Our work is focused especially on the pitch frequency and energy parameters.

Lip shape based emotion identification

Emotion recognition systems have an important role to play in the human-computer interactive applications (HCI). These systems are using facial features of face images and they are verifying or identifying the emotions. In this study, emotion identification algorithms are improved by using just mouth region features of a face. Region of interest (mouth region) is detected by Viola-Jones algorithms from video frames which are including different emotional face expressions. Outer boundaries of lip shapes are extracted by manually and calculated the scalar Fourier Descriptors (FDs) of the boundaries. Classification and recognition of the emotions is presented according to scalar FDs of lip contours. Test results are obtained as 93.9 % accuracy rate for scalar FDs.

Segmentation of isolated words into voiced-unvoiced sound components by kurtosis

This study presents a new approach to the segmentation of isolated words into their voiced/ unvoiced parts. It is well known that voiced/ unvoiced discrimination has an important role in speech synthesis and coding applications. The offered method makes this discrimination using the kurtosis values of the words. The performance of the proposed approach was tested on Turkish digit recordings from zero to nine. It has been observed that this approach segments the parts successfully in not only clean speech but also in noisy speech.

Recognition of Hand-Sketched digital logic gates

Hand-Sketched circuit recognition is a very useful tool in engineering area. Because most of the engineers prefer to design their circuits on the paper firstly. So, this can cause time wasting and some mistakes. In this study, which is based on the solving these kinds of problems, classification and recognition of the handwritten digital logic gates according to their complex and scalar FDs (Fourier Descriptors) is presented. Test results are obtained as 84.3 % accuracy rate for complex FDs, 98.6 % for scalar FDs. Then these results are compared and decided the optimum FDs type for this study.