H. Gokhan Ilk

Priority and significance analysis of selecting threshold values in Adaptive Unsharp Masking for infrared images

Infrared image processing has been the focal point of considerable research activity in the last decade mainly because of its wide application areas in security and defense. With the aid of an existing image enhancement technique, Adaptive Unsharp Masking (AUM), proposed optimum parameters selection procedure delivers better performance in sharpness and contrast adjustment for the detection of targets in interest in objective quality metrics. In our study, we focus on the lower and upper limit threshold parameters which classify the contrast areas and therefore have significant effect on sharpening the edges for two different infrared (IR) images. The experimental results prove that changes of convergence rate of adaptive filter and positive convergence parameter versus mean square error (MSE) exhibit similar characteristics under certain lower and upper threshold values which satisfy “minimum” MSE; that is these threshold values are the most dominant criterion amongst the other parameters and they are significant.

Fujisaki intonation model in Turkish Text-to-Speech Synthesis

In this work, a Turkish Text-to-Speech Sysnthesis system has been developed. The system is a concatenative one using diphones. Then, Fujisaki intonation model [1], which is known for its successful results in many languages is applied to the system for increasing the naturalness of the synthesis. The F0 contour of the synthesis resembles the original very much and the results seem promising. There has been an increase of 0.15 points in the PESQ result of a sample sentence to which the intonation model has been applied.

Automatic license plate recognition based on a projection method

Automatic license plate recognition system is used numerous applications such as: red light enforcement, parking entrance, highway toll fee collection ect. In this study automatic license plate recognition system is presented in a different method. First, Mean Value Algorithm is used in order to locate the plate. Then the characters of the plate are segmented. Finally, the characters are recognized by using a correlation method. The system is tested by a database which includes 128 images of different vehicles where the success rate varies 70%–97% according to the parts of the system.

Human action recognition via spatial and temporal methods

In this work human action recognition problem was discussed in video sequences. Solution of the problem was studied in three stages. Firstly, points of interest were detected with preproccesing and these points which are called cuboids were declared in small windows, then feature extraction was performed and finally, human action is decided by using classification. Features extraction is not only performed in the spatial domain but also along the cuboid video, that is in the time domain. K-nn(nearest neighbor) algorithm was used as a classifier. Furthermore, algorithm was run on the Weizmann database and results were presented. In addition to the traditional human action studies, different databases were evaluated in this work, preprocessing and feature extraction parameter optimization was also performed. The results show that performance is increased by 6-11%.

Speech coder identification using false nearest neighbor fractions

Speech coder identification refers to identification or recognition of a speech coder directly from a bit stream. The fundamental idea behind coder identification is that different coders should posses different statistical bit patterns and therefore recognition should be possible using chaotic features such as false nearest neighbor fractions. In this study theoretical coders have been investigated first and then research has been extended to real time actual speech coders. The results are impressive, in the sense that the proposed speech coder identification system that operates on nine different conventional coders delivers between 96.36 to 100% performance.

Syndrome code data hiding using statistical modeling with Markov chains

Some fields of an encoded speech or audio signals bit stream, which varies according to encoders type, can be modeled by Markov chains. In this paper, a novel “syndrome code data hiding using Markov chains” is proposed where secret data is embedded into the syndromes of the C(N, K) linear block codes as in matrix embedding. However, the proposed method randomly chooses a codeword from 2K possible code words according to the Markov chains transition probabilities, which is different from matrix embedding method. Performance of the proposed method is compared with that of least significant bit and matrix embedding methods employed on GSM 6.10 coder. The simulation results show that data hiding using statistical modeling with Markov chains preserves the original bit streams entropy, leading to undetectability in terms of steganalysis. Unfortunately secret data embedding efficiency is decreased.

Steganalysis of the MELP quantization index modulation data hiding methods by using chaotic type features

The success of data hiding methods, which hide data to media like speech signal, is primarily determined by their strength aganist to the steganalysis methods. In this paper, some data hiding methods, which embed secret data during MELP analysis of speech signal using quantization index modulation are introduced. These methods in question are examined by a steganalysis method which takes advantage of chaotic-type features of speech signal. Thus, by evaluating the steganalysis results, the practical usage limitations of data hiding methods can be exposed.