Browsing by Author "TORA, Hakan"
Now showing 1 - 4 of 4
Results Per Page
Sort Options
Item A NEW APPROACH FOR SELF-CALIBRATING CAMERAS(2022-02-16) Gurel, Cahit; TORA, Hakan; GÜNEŞ, AhmetCamera is one of the most important sensors in robotic applications. Calibrated ca meras provide more information than the uncalibrated ones. Intrinsic parameters of a camera can deteriorate due to mechanical and thermal changes in environment. There fore self-calibration is required for robotic operations. Since self-calibration does not require any known template objects in the process, it is more flexible and extract ing a few fixed points between calibration images is enough for self-calibration. We propose a new method for simpler and more accurate self-calibration method by in corporating some of the extrinsic parameters of camera along with some assumptions which are true for present day cameras. Moreover, we have included a basic point de tection, tracking and association approach for the task. Proposed method is tested and compared with another self calibration method using synthetic data, a mobile robot with a camera in V-REP simulation environment and physical implementation with articulated robot arm. The results indicate the effectiveness of the new approach with respect to other self-calibration approaches for planer motion of the camera.thesis .listelement.badge DESIGN AND IMPLEMENTATION OF SPEECH TO SIGN LANGUAGE TRANSLATOR SYSTEM(2022-01-10) TAMEEMI, AHMED HASHIM HAMZA; TORA, HakanNumerous efforts have been to design interpretation strategies between the healthy and deaf people. One of the most important of these systems is what called as Human-Computer-Interaction (HCI) systems. Many simulation works of sign language are evolved in the last years in this field. At most, these efforts have been worked to convert Sign Language into English speech and text, and verse versa. Usually there is minimal interaction between healthy persons and the Arabic deaf persons. So it is very important to design a conversion system that can translate ArSL into Arabic speech or text and vice versa to provide a better communication between normal persons and Arab deaf community. In this thesis, a translation system for Speech to Sign language will be accomplished; the target is to design a translator scheme that can translate the speech signals of Arabic letters into display the standard Arabic sign language. This circuit can be used to realize a human-friendly program scheme that can be used for numerous applications such as communication between the deaf and the normal people. Several stages have been carried out to gratify the system design requirements, including: recording the signals of speech for data collection, feature extraction after pre-processing, and the recognition final step. Then, the system provides the recognition factors to the translator structure, to display image of movements of fingers that antipodean the spoken letter. In this work, Pattern Recognition Neural Network (PRNN) was used. It is a feedforward network that has an ability of classifying input samples into desired classes. The network was trained with back-propagation algorithm by 560 training samples, 20 samples for each letter of the entire 28 main Arabic letters. iii Then, the trained model was validated and tested with 140 and 140 dataset respectively, to see how fit the PRNN model predicts the matching data set of output labels. The network is successfully trained for all 28 classes (letters). The recognition was achieved with an excellent diagnosis rate of up to 98%. ÖZET: Sağlıklı ve sağır insanlar arasında yorumlama stratejileri tasarlamak için çok sayıda çaba sarf edilmiştir. Bu sistemlerin en önemlilerinden biri İnsan-Bilgisayar-Etkileşim (HCI) sistemleri olarak adlandırılan sistemlerdir. Bu alanda son yıllarda birçok işaret dili simülasyon çalışması geliştirilmektedir. Bu çabalar en fazla İşaret Dili'ni İngilizce konuşma ve metne, ayeti ise İngilizceye dönüştürmek için çalışıldı. Genellikle sağlıklı kişilerle Arap sağırları arasında minimum etkileşim vardır. Bu nedenle, normal kişiler ve Arap sağır topluluğu arasında daha iyi bir iletişim sağlamak için ArSL'yi Arapça konuşmaya veya metne çevirebilen bir dönüştürme sistemi tasarlamak çok önemlidir. Bu tezde, Konuşmadan İşaret diline çeviri sistemi yapılacaktır; hedef, Arap harflerinin konuşma sinyallerini standart Arap işaret diline çevirebilen bir çevirmen şeması tasarlamaktır. Bu devre, sağırlar ve normal insanlar arasındaki iletişim gibi çok sayıda uygulama için kullanılabilecek insan dostu bir program şemasını gerçekleştirmek için kullanılabilir. Sistem tasarım gereksinimlerini karşılamak için, veri toplama için konuşma sinyallerinin kaydedilmesi, ön işlemeden sonra özellik çıkarımı ve tanıma son adımı dahil olmak üzere çeşitli aşamalar gerçekleştirilmiştir. Daha sonra sistem, konuşulan harfin tersi olan parmak hareketlerinin görüntüsünü göstermek için çevirmen yapısına tanıma faktörlerini sağlar. Bu çalışmada Örüntü Tanıma Sinir Ağı (PRNN) kullanılmıştır. Girdi örneklerini istenen sınıflara ayırma yeteneğine sahip ileri beslemeli bir ağdır. Ağ, toplam 28 ana Arap harfinin her harfi için 20 örnek olmak üzere 560 eğitim örneği ile geri yayılım algoritması ile eğitilmiştir. Ardından, PRNN modelinin eşleşen çıktı etiketlerini ne kadar uygun tahmin ettiğini görmek için eğitilen model sırasıyla 140 ve 140 veri seti ile doğrulandı ve test edildi. Ağ, 28 sınıfın (harflerin) tamamı için başarıyla eğitilmiştir. Tanıma, %98'e varan mükemmel bir teşhis oranıyla sağlandı.Item IMPLEMENTATION OF TURKISH TEXT TO SPEECH SYNTHESIS WITH RC8660 VOICE SYNTHESIZER(2015-10-25) KARAMEHMET, Timur; TORA, Hakan; USLU, İbrahim BaranThis thesis examines the text-to-speech (TTS) synthesis problem and the adaptation of the RC8660 Embedded System to Turkish. RC8660 is a system loaded with English phonemes and in the thesis the aim is to make this card synthesize Turkish. Firstly Turkish phonemes from their corresponding English phonemes are defined. For this purpose IPA: International Phonetic Alphabet was used. Because the syllabic structures of Turkish and English are different, there occurred a need for defining an Exception Dictionary using the text and phoneme modes of the system. The rules for correct syllabification were added to the dictionary one by one. Stress and intonation rules were also defined for making the synthesized Turkish speech as natural as a native speaker’s. Different phonemes were used for text and alphanumeric characters. The effects of the Speed, Expression, Pitch, Formant Frequency, Tone, Delay, and Articulation adjustments in the system’s software: RC Studio were also tested on the speech. The quality of the synthesis was evaluated by the Mean Opinion Score (MOS) test.Item SPEAKER INDEPENDENT ISOLATED DIGIT RECOGNITION(2022-01-20) Hamid, Mohammed Saeed Hamid; TORA, HakanIn several speech signal processing applications, VAD presents an important character for splitting an audio stream into time intervals that include speech activity and time intervals where speech is absent. In this research, we presented new approach dealing with isolated word recognition. In the first stage, three functions applied for voice activity detection (VAD) problem hamming window, Bohman function, and Bartlett-Hann function. The both Bohman function and Bartlett-Hann function are not applied in previous studies for VAD problem. On the other hand, pitch, MFCCs, and energy applied as feature extraction techniques and combined with SOFTMAX which these two methods are new approaches. The Pitch based SOFTMAX presented remarkable results which extracted features by pitch wired to SOFTMAX and classified to seven words and presented 85% accuracy. Furthermore, energy also applied as feature extraction and the output of this function wired to the SOFTMAX. This framework easily can applied to the various isolated word recognition which only the user modified the input data easily. The main contribution in this study, combine SOFTMAX with several feature extraction techniques. The SOFTMAX is trend probability function which analysis input features to the labels between (0,1) and used in several deep learning techniques as last layer function for classification or regression issues. The obtained results compared with several studies presented in this field by applying several machine learning and deep learning techniques combined with audio signal processing techniques that’s applied for feature extraction.