Theses supervised by Prof. Dr. Ergun Erçelebi
68 theses · Gaziantep University
Development of maximum power point tracking system (mppt) for solar panel
Due to the global energy crisis and environmental problems, the utilization of photovoltaic (PV) systems as an electric source has increased exponentially. It is expected that this usage trend will keep increasing in the near future too. However, the output power from the photovoltaic modules is affected by radiation intensity of the solar cells. Therefore, to exploit all power of photovoltaic panel and maximize the efficiency of the photovoltaic (PV) systems, it is necessary to track the maximum power point of the input source. In the thesis, the efficiency of perturb and observe (P and O) maximum power point tracker (MPPT) has been enhanced. The efficiency improving has been achieved by ameliorating the tracking speed (transient response) and decreasing the steady state oscillation. A step-size perturbation is utilized to decrease the oscillation and minimize the reaching time of operating point to MPPT. When the operating point is far from the MPPT, the perturbation is made larger to ensure that the tracking speed is maximized. But as the operating point gets closer to the MPPT, the perturbation reduces. The algorithm proposed for maximum power point tracking system (MPPT) for solar panel has been realized on boost converter by design of electronic card having STM32 microcontroller. The card scans the gained power of PV panel over short periods to adapt the nonlinear electrical characteristics of PV panel. Extensive experimental and simulation results have shown that the proposed algorithm is effective in variable weather conditions.
Development of multi-level inverter for power quality enhancement
The objective of this thesis is to develop and improve the performance of the multi-level inverter. Two paths have been followed in the development and improvement works. The first one represents the off grid connected multi-level inverter. The trend is focused to increase the number of the level at the output of the inverter while the inverter represents the main energy source. In the second path followed in the studies, the grid-connected multi-level inverter has been considered. Power quality enhancement has been done by developing the control unit of grid connected multi-level inverter in such a way to ensure the stability of the system against the system voltage sag. The multilevel inverters designed in the thesis are both simulated using a computer software program and implemented in hardware. The developed inverters are multilevel and hybrid inverters. Conventional inverters give five levels in the output waveform. These inverters contain two H-bridge circuits to generate different levels at the output while the hybrid inverter developed in the thesis includes three H-bridge circuits to form different levels at the output. Moreover, developed hybrid inverter can also operate in conventional mode when it is supplied with equal sources. The output of each H-bridge has been connected in a cascaded form. The sum of the individual H bridges voltages represents the final inverter voltage. Experimental results on the developed circuits gave satisfactory results and confirmed the accuracy of the design. Key Words: Low Voltage Ride Through (LVRT), Photovoltaic Panel (PV),Multi Level Inverter, Maximum Power Point Tracking (MPPT), Active And Reactive Power Control
Development of RW-UAV detection model using deep learning training with a synthetic dataset
The use of rotary wing unmanned aerial vehicles (UAV) is increasing day by day and their accessibility is getting easier. As a disadvantage of easy accessibility and usability, UAV technology threatens security and personal areas. This technology, which is easy to access, should be detected with a cheap and easy-to-install detector. The aim of the study is to develop the UAV detection sensor using camera and machine learning based systems. Deep learning method has been used in sensor technology, which has shown successful classification results recently. Artificial intelligence trained with deep learning method shows effective results in terms of performance. On the other hand, artificial intelligence (AI) training needs a lot of data. To overcome this obstacle, the data set needed by the deep learning algorithm was produced synthetically using game engine software. The original aspect of the study is that the artificial intelligence trained with synthetically produced images was tested on real data and achieved 90% success. The success of AI, which is directly trained with synthetic data and tested with real data, is very low. The reason for this is the difference between the synthetic and real image. To minimize this difference, 7 different experiments were designed for different feature extraction algorithms. One of the experiments is the Corner Detection and Nearest Three-point Selection (CDNTS) feature extraction layer that we originally developed. The CDNTS layer classified the real data with 90% success.
Design and analysis of wearable antennas for use in wireless biomedical telemetry systems
This thesis presents the design of a fractal printed circuit antenna construction for portable applications. A modified Hilbert patch was printed on a flexible polyamide substrate for the proposed antenna. An open loop with a defective electromagnetic band gap (EBG) surrounds the antenna. Experimental findings are compared to simulated results from a commercial software package based on Computer Simulation Technology Microwave Studio for the proposed antenna (CST MW). The experimental and simulated findings are found to be quite similar. The suggested antenna design is determined to be compatible with the frequency ranges of UHF and ISM portable systems. The antenna shows two bandwidths from 0.8GHz to 1GHz and from 2.3GHz to 2.6GHz with average gain of around 3.1dBi and 7dBi, respectively. The antenna performances are conducted based on analytical and numerical simulations. The fabricated prototype performances have been measured experimentally to show excellent agreement with the simulated results. Finally, it is found that the proposed antenna frequency resonance and bandwidth are insignificantly affected when mounted in close to the human body; therefore, it would be excellent candidate to be in use with biomedical applications. The proposed antenna shows several resonances around the Body Area Networks (BAN) and ISM frequency bands. The proposed antenna shows a suitable gain for short and medium wireless communication systems about 1dBi, 1.24dBi, 1.48dBi, 2.05dBi, and 4.11dBi at 403MHz, 433MH, 611Mz, 912MHz, and 2.45GHz, respectively. The antenna is printed using silver nanoparticle ink on a polymer substrate. The antenna size is reduced to 20×10mm2 to suit the different miniaturized wireless biomedical devices. The fabricated prototype has been tested experimentally on the human body. The main novelty with this design is suppressing the surface wave from the patch edges, allowing it to significantly reduce the back radiation toward the human body when used enclosed to it.The proposed antenna is located on the human head to determine the specific absorption rate (SAR). It is found in all cases, the proposed antenna shows low SAR effects on the human body.
4 wheel steering electronic control unit design
Ease of driving and flexibility in movements are very important for the developing automobile technology today. The aim of the study is to provide convenience to the driver by trying different modes in the steering system of a vehicle. In the thesis, an electronic board has been designed for the steering system for the 4 wheels in the vehicle. The card developed in the study provides choosing from 3 different modes as standard, circle and crab to driver. The driver will have the advantage of both comfort and safe driving with the mode chosen according to the road and load conditions. Also, the 4-wheel steering system improves handling stability and active safety at high speeds. Using 4-wheel steering at low speeds can reduce the turning radius on the steering wheel for more convenient maneuvers. Positive feedback was received from drivers who tried the card on heavy-duty vehicles. Tests continue on different heavy-duty vehicles. If the electronic board and software do not perform as expected in different heavy duty vehicles, improvements will be made in the design. However, the results obtained so far are satisfactory. It shows good response performances, also improves ride quality and stability
Building management system for distribution control center
As the complexity and the type of electrical applications perpetually expands, monitor and control will require new systems are based on some features in such feature of self-powered and distantly management. Building management systems (BMSs) are developed as efficient solutions for users to controls and monitor electronic devices. BMS is a method especially developed to monitor environmental events such as temperature, fire and humidity in the computer server room and to warn the operator in case of fire or if the temperature and humidity values exceed the specified threshold level. A BMS uses some inexpensive sensors and real-time modules to generate advance information about the climate inside the space that needs to be remotely controlled. In this thesis, a building management system for a computer server room has been developed using inexpensive sensors, prototyping cards and mini-computers. Raspberry Pi used as a mini computer has been utilized to collect information from sensors in the server room or data center and transfer them to the receiving side via ZigBee wireless modules. The Arduino prototyping card has been used to process the sensor data transmitted wirelessly from the server room and actuate the actuators such as ventilation fans and switches. The developed BMS has been tested under critical environmental conditions to evaluate its performance. Experimental results showed that the BMS performs robustly and precisely even in extremely unwanted conditions
A real-time face recognition based on mobileNetV2 model
Facial recognition technology is one of the fastest developing technologies. Thanks to its efficacy, ease of use, and extensive applications in health, protection, and surveillance, face recognition is the most widespread technology compared to other biometric ones. Technologies based on deep learning and neural networks have demonstrated superior efficiency and speed when compared to traditional approaches for recognizing persons. In this work, we propose a fast real-time facial recognition system that is applicable for mobile devices and devices of low computational power. It relies on today's latest convolution neural networks algorithms. The database is built based on photos from a collection of known people and some VggFace dataset celebrities. The proposed system is divided into several steps, starting with the detection of faces in the input images using the MTCNN algorithm, followed by their alignment and preprocessing, extracting the face characteristic vectors for each face using the mobileNetV2 model, and finally comparing, classifying and distinguishing the faces. The performance of the system has been evaluated using some examples. The system produced results with an accuracy of 92.67%, with an average of 11.68 frames per second. When compared to the state-of-art models, the experimental results demonstrated that mobilenetv2 model was 2 times faster than ResNet50 model and 4 times faster than VGG16.
Pre-surgical device for correction of the cleft lip
The first object of this thesis is to propose a new architecture that can be utilized in the pre-surgical treatment of cleft lip and palate. The architecture consists of two main parts mechanical and electronic. The proposed architecture provides an external pressure applied directly to the cleft lip areas through an airbag placed on the mechanical part of the architecture. The amount of pressure applied to the cleft lip is controlled by the electronic part. FEA method was used to analyze the effect of pressure applied to the cleft lip and cleft bone regions. We analyzed the effect of displacement distributions, the tensile and compressive strengths, and von Mises stress distributions on the cleft areas. The results proved that the external pressure led to bringing the incisions closer to each other. External pressure has contributed to bringing the incisions closer to the target position of the plastic surgeons. The results have shown that the proposed architecture can contribute to the presurgical treatment of clefts in the mouth and palate as an alternative to the traditional method. The second object of our thesis is to detect the cleft lip and extract its features that provide a second opinion for plastic surgeons for a vision of the cleft lip features using the combined method. The combined method consists of three algorithms for cleft lip detection and feature extraction (face detection, cleft lip classification, and cleft lip detection). The results show that our method has the best efficiency and high accuracy in lip detection and optimal feature extraction. Moreover, the proposed method demonstrates its effectiveness with a limited set of cleft lip images.
COVID-19 detection from signs and symptoms using machine learning
Effective SARS-CoV-2 (COVID-19) screening enables a speedy and precise diagnosis of COVID-19, lowering the load on the health care systems. Different machine learning models were developed for COVID-19 detection. These models are intended to aid physicians globally in the quarantining of patients, especially in poor areas. At the end of 2021, the global officially recorded COVID1-19 cases reached 286,582,541 with 5,430,949 confirmed deaths. This study aimed to deploy and test machine learning-based models for predicting COVID-19 diagnosis. Seven machine learning models have been utilized as Naïve Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), Multilayer Perceptron Neural Network (MLP), Decision Tree (DT), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). The models were trained and tested on data from 2,151,898 tested people, among whom 208,726 (9.7%) were found to have COVID-19. Simply six binary characteristics were used to predict COVID-19 testing results with acceptable accuracy. The data included five early COVID 19 clinical signs and symptoms, such as cough, fever, sore throat, shortness of breath, and headache. Also, the data from people in contact with a confirmed COVID-19 case were used. Overall, we evaluated the possibility of building a model that could be used for prioritizing testing for COVID-19 by asking simple questions depending on the data we acquired from GitHub. When using normal unbalanced data; we achieved 92.9% accuracy, 60.4% sensitivity, 96.4% specificity and 62.2% F-score with the MLP, DT, RF and XGBoost machine learning models. When using balanced classes, all the seven machine learning models produced 91.8% accuracy, 65.1% sensitivity, 94.7% specificity and 60.61% F-score results. When using only symptomatic cases (161.915), MLP, DT, RF and XGBoost models gave 70.8% accuracy, 89.1% sensitivity, 47.3% specificity, and 77.4% F-score results. In the three scenarios XGBoost, RF, DT, and MLP gave the best results. Regarding the training time, XGboost, RF and DT were faster than MLP. Key Words: Machine Learning, Classification, Diagnosis, COVID-19, SARS-COV-2, Signs and Symptoms
Design of frequency selective surface based on reflection coefficients in C and X bands using nested convolutional neural network
Frequency selective surfaces (FSS) are surfaces consisting of periodically placed shapes used to reflect, transmit and absorb electromagnetic waves. FSS has many uses such as increasing the antenna performance, reducing the radar cross section, filtering, and increasing the photovoltaic cell performance. Due to the use of FSS in many fields, it has attracted the attention of researchers and has led to studies in this field. The studies are generally aimed at obtaining the FSS surface with optimization algorithms and methods. For this purpose, the desired reflection coefficients (S11) are obtained by iteration. With the significant increase in computational power in recent years, the use of machine learning algorithms has become more widespread and successful results have been obtained in classification and regression problems. In this thesis, studies were carried out to obtain the FSS corresponding to the desired S11 with Decision trees, K-nearest neighbor, Deep learning and hybrid deep learning algorithms from machine learning algorithms. In the training process of machine learning algorithms, reflection coefficients between 6-14 GHz were used as input, while 50x50 FSS images were used as targets. Machine learning models were trained using 29722 reflection coefficients in total. As a result of the training process, the most successful result was obtained with the nested convolutional neural network model, which is a hybrid model. The mean absolute error value of the nested convolutional neural network model was 0.69, and the R2 value was 0.9.
Raspberry pi ve python ile gerçek zamanli plaka tanima sistemi
In order to provide reliable transportation with the increasing traffic density in the current time, it is necessary to detect dangerous situations with video analysis by using camera and remote control system. One such analysis is Automatic License Plate Recognition. Automatic license plate recognition systems consist of algorithms such as license plate detection, edge detection, character segmentation and optical character recognition. With the license plate recognition system developed in this thesis, after preprocessing the license plate image taken from the camera with image processing algorithms, license plate recognition has been made with the deep learning algorithm and the recognized license plate information has been sent to the defined e-mail address. In the developed license plate recognition system, Raspberry Pi, called mini computer, has been utilized as hardware, and the software has been developed in Python programming language. From edge detection methods; the performances of Canny, Fuzzy, Roberts, Sobel and Prewitt methods on the license plate recognition system were measured and the performance of the Canny edge detection method was found to be high and the algorithm of this method was embedded in the license plate recognition system. As a final step, Tesseract optical character recognition has been performed by scaling the frame. After recognition, a comparison has been made with the characters defined in the deep learning structure and the license plate recognition process has been completed. Tests were carried out to evaluate the developed license plate recognition system. As a result of these tests, the average accuracy rate of the system was 91.67% in real-time license plate recognition and edge detection, and 89.43% in character recognition. With the developed license plate recognition system, low cost and high accuracy recognition will be achieved.
Heart attack prediction using machine learning
According to the World Health Organization (WHO) 2020 report, more than 17.9 million people lose their lives due to heart attacks each year, making it one of the deadliest diseases worldwide; furthermore, heart attack is classified as a cardiovascular disease that affects the lives of many people globally. Machine learning, vital to the healthcare industry, enhances people's quality of life along with scientific advancements. In this thesis, studies on predicting heart attacks using machine learning algorithms were conducted. The data used in the studies were obtained from the UCI Machine Learning database. The prediction of heart attacks was carried out using five well-known machine learning models mentioned in the literature. Additionally, a new conceptual model integrating data preprocessing, hyperparameter tuning, and Logistic Regression was proposed for predicting heart attacks in this thesis. Tests conducted on the proposed model yielded satisfactory results, achieving an accuracy rate of 93%. Furthermore, the results produced by the proposed model were compared with the outcomes of other machine learning algorithms, demonstrating that the proposed model outperforms other algorithms.
Face detection and recognition based on raspberry Pi using HAAR cascading and convolution neural network
Face Detection is a form of biometric method that relates to the automatic detection of faces by computerized systems through observation of the face. It is a popular feature in biometrics, digital cameras, and social tagging. Face detection and recognition have received increased research focus in recent years. In this thesis, a face detection and recognition system have been proposed and developed for detecting and recognition faces through the hybridization of two algorithms: HAAR cascading algorithm and deep learning algorithm. The proposed system consists of two approaches. The first approach, the HAAR cascading algorithm, was developed by taking a shot of the face and reducing it several times to ensure that there is a face at each shrinking time. The second approach has proposed convolution neural network (CNN) model to increase accuracy of classification. In addition to improving each algorithm, hybridization of the two algorithms significantly improved the results of the classification. In proposed system two dataset was used: download dataset, and real dataset. The accuracy of modifying HAAR in detection reached 98.667% for real dataset, and 97.532 % for download dataset. The accuracy of proposed model of CNN in classification reached 96.23% for download dataset, and 100% for real dataset. The tests conducted on the developed facial recognition system have demonstrated that the proposed algorithms and the developed real-time facial recognition system yield satisfactory results. Key Words: Face Detection, Face Recognition, HAAR Cascading, CNN.
Development of 3d laser scanner using laser range finder
Generally, 3D laser scanners are used to create 3D model file of any object or any environment to help architecture or robot navigation. A 3D model can be generated using one camera and laser light source or stereo camera or a laser distance meter. In this study, Laser scanner hardware has been designed and implemented to create a 3D model in AutoCAD software using a laser range finder. The hardware can generate 3D modeling of environment using the distance measures from different points on spherical coordinate plane. Two axis (φ, θ) movable mechanical parts have been designed and implemented to get distance measurement from different points. Step motor driver has been developed to control the step motors that provide the movement of mechanical parts. Another electronic card has been developed to provide duplex communication between computer and electronic device. A computer software and GUI have been coded that are able to generate modeling using collected data. Approximate modeling of environment can be generated by computer software owing to data that are collected by scanner. These datasets contain necessary distance information to detect borders, barriers or proper ways for a robotic system that roams environment. A laser scanner can obtain information about distance between objects in environment. Some information about environment like slope of ground or object distribution can be analyzed by these datasets
Design of channel estimation and equalization for OFDM systems
Orthogonal Frequency Division Multiplexing (OFDM) system is one of the multicarrier techniques which is robust against Inter-symbol-Interference, multipath fading and very easy to apply in transmitters by using inverse fast Fourier transform IFFT and at the receivers by using fast Fourier transform FFT. In a communication system, channel estimation is very important issue for the data detection. In coherent detection, one of the popular techniques is to use pilot tones as a reference signal in OFDM symbols. In the comb-type pilot tones insertion, pilot tones are inserted into each OFDM symbols, but inserting a large number of pilot tones will lead to channel capacity reduction or bandwidth expansion [1-2]. In this thesis, to overcome this transmission loss, a modified least square (ModLS) algorithm for fast time varying wireless channel at comb-type pilot arrangement in QAM signals for OFDM system is proposed. The proposed algorithm has reduced number of pilot tones in each OFDM symbol. 1/100 pilot ratio has been achieved. This ratio ranges from 1/16 to 1/18 in other studies. Also, we have improved bit error rate (BER) in low signal to noise ratio (SNR) by the proposed algorithm. The simulation results obtained from the proposed algorithm showed a good performance in noisy wireless channels. In addition, it has been compared with least square (LS) algorithm in different signal to noise ratios and different channel tabs. The comparison results showed that the proposed algorithm was superior to least square algorithm.
Application of monomodal intensity-based medical image registration on brain tumor growthiness investigation
The process of aligning two medical images (MR/MR, MR/CT, MR/PET, CT/PET etc.) and taking those medical images to the same coordinate system is called medical image registration. The images can be acquired with different sensors (MR, CT, PET, SPECT etc.) at the same time or with the same sensor at the different times. Registration concept is one of the most important and popular aspect of digital image processing. Thanks to using suitable computer programming techniques and transformation between two images a new much more informative image can be found. In this thesis, monomodal registration of MR image of a patient with brain tumor with another MR image of the same patient at a different time is proposed so that the growthiness of the tumor inside the brain can be investigated. Registration is carried out using intensity based technique and Sum of Squared Differences (SSD) metric with the regular step gradient descent optimizer. The advantage of this method is that there is no need to calculate image features such as Fourier transform, eigenvalues and eigenvectors. This technique is superior to other techniques such as cross-correlation method and mutual information similarity measures because it avoids the wrong registration and uses less computation.
Fast encoder implementation in high efficiency video coding (H.265/HEVC)
The rapid development and increase of multimedia applications, as well as the demand for higher video-quality services at restricted resources like storage capacity, transmission bandwidth and power consumption, has brought the urgent need for more efficient video compression techniques. The emerging video coding standard which has recently standardized as High Efficiency Video Coding (H.265/ HEVC) has a significant superiority over its predecessor, Advanced Video Coding (AVC/H.264). When compared with AVC, HEVC is reported to halve the bit-rate with the same visual quality, or a better quality with the same bit-rate. Beside other improvements, HEVC significantly gets its power from the use of dynamic hierarchical quad-tree structure by partitioning the frames into smaller regions called coding units (CU), by means of a rate-distortion optimization process (RDO). However, this improvement yields to a dramatic increase of high computational complexity and increased encoding time, which primarily restricts its adaptation in real-time applications. This thesis briefly represents the structure and principles of HEVC and evaluates its outperforming performance against previous video coding standards, including AVC. In our study, an early CU determination algorithm for fast encoder realization is proposed in order to reduce the encoding time which is the most important part of the standard standing for development.
Design and implementation of DC/AC converter for photovoltaic panels
An inverter is a major component of solar photovoltaic (PV) system for combining photo voltaic and battery. In this study, 600W high efficiency (%95) advanced inverter relevant to PV system has been designed and implemented. The inverter consist of two parts; the first is the synchronous DC-DC boost converter with variable input voltage from 9V to 60V and 12 to 80V DC constant output voltage according to the set point. The second part is DC-AC voltage source inverter. The sinusoidal bipolar SPWM has been used for triggering the gates of the IGBT switches in the inverter bridge. dsPIC16F877A microcontroller as embedded system part of inverter has been utilized. The MATLAB, Simulink has been used for simulation studies on the inverter. Experimental results and simulation studies show that the performance of our design is satisfactory.
Comparison of techniques for voltage control and power system stability in the north of Iraq
Power flow and voltage control are a vital role in the north of Iraq's electrical power system. This thesis illustrates the potential applications of flexible AC transmission system (FACTS) controllers, such as the static VAR compensator (SVC) and the Static synchronous compensator (STATCOM) for the control of voltage and power flow, and improving the voltage regulation. The STATCOM and SVC are used in 132 kV Soran substation. Based on, a back to back connected thyristor valves and voltage-source converter, SVC and STATCOM regulate system voltage by generating or absorbing reactive power. The objective is to define the reactive power generated and voltage control at 132 kV Soran substation using STATCOM and SVC. Finally, in case, our study has compared the STATCOM with SVC in static voltage stability improvement. Different performance measures have been considered under different operating system conditions for the north of Iraq's electrical power system. The results show that STATCOM has better performance over SVC.
Smart power system management and control of water stations of Garmian region in Iraq by using micro grid
In this thesis, we have developed a power management system to optimize the power of water station in Garmian (Iraq). A novel approach has been proposed for control of the flow of power in the micro-grid network. The proposed approach includes the typical five microgrid key elements such as micro sources (energy sources), loads, storage devices, and control elements, and point of common coupling. The goal of designing such a system is to provide interruptible high-quality power to sensitive loads in assured area. A model for this goal has been developed and simulated by MATLAB/Simulink programming. Also, smart power management of water station has been designed by using the external controller for Garmian region in Iraq.
Modeling and implemention of variable frequncy drive based on open-loop volt/hertz
A single or three -phase squirrel cage motors are one of mostly used in many industrial applications this motors consuming more than half of the total generated electrical energy .The Induction motors can run only at their full speed when they are attached to the AC supply. However, speed control of squirrel cage motors is important to many industrial applications that require variable flow control of fluid (air, water and chemical liquid). Speed control method that has been developed with voltage-source inverter become a popular method for the speed control of induction machines. The purposes of this thesis are to model and implement an AC-Drive, by three type inverter switching techniques as Sinusoidal Pulse width Modulation (SPWM), Third Harmonic Injection Pulse Width Modulation (THIPWM), and Space Vector Pulse Width Modulation (SVPWM). These techniques are based on Voltage-Source Inverter (VSI) and open- loop speed controller, with principle (Volt/Hertz). The three switching techniques and open-loop (Volt/Hertz) control are compared with each other's. The mathematical model has been developed for AC-drive system. Also, the model has been simulated by a software to evaluate the performance of open-loop speed control system with three switching techniques. The results indicate that the mathematical model is accurate and practicable, also achieved to satisfy the requirements of the users. Moreover, in the thesis, a practical model of AC-Drive with SPWM technique has been realized. The practical model imitates to conceptions of power electronic module, (HY-DQ Inverter v 2.0) module used to control and drive of the three phase squirrel cage induction motor. This practical model has quality of completely implemented and exact results are obtained. The experimental results are based on the voltage measured in load side and rotor speed with actual load (1-hp IM). These results have shown satisfactory performance of the system from output voltage waveform and rotor speed.
Development of counterfeit banknote recognition system using deep learning
Banknote counterfeiting undermines financial stability and public trust on the currency of a nation. Moreover, it also has various economic implications on financial institutions and businesses. The counterfeiters employ increasingly sophisticated techniques to mimic the genuine banknotes with the advancement of technology, which necessitates the development of fast and robust counterfeit banknote detection systems in order to ensure the integrity of financial systems as well as maintaining the public trust on currency transactions. In this thesis, a Turkish lira authentication system has been developed which locates the banknotes on images and classify the located banknotes as genuine or counterfeit as well as identifying their denominations. The localization part is developed using the classical and well-known digital image processing techniques, where a novel convolutional neural network architecture is developed for the classification part as the deep learning, which is a subset of artificial intelligence, has shown remarkable success in automated feature extraction and pattern recognition tasks. The conducted experiments shows that the proposed architecture outperforms some state-of-the-art deep learning algorithms in terms of accuracy, loss, inference speed, and file size.
Spectrum sensing and opportunistic medium access control protocol design for cognitive radio wireless networks
The explosive growths in the wireless communication field require more spectral resources, while the radio spectrum is a finite natural source. On the other hand, usable frequency bands are nearly fully occupied by the licensed users by the static allocation strategy and it is also known that most of the radio spectrum is inefficiently used. In cognitive radio (CR), frequency allocation and the access to radio spectrum is not assumed fixed but dynamic and thus cognitive radio technology utilizes this dynamic structure to improve spectrum efficiency and network performance. The primary objective of the thesis is to study spectrum sensing duty of cognitive radio attached with proper medium access control (MAC) design. A new wireless network scheme where the intentional background noise is added to the wireless region is proposed. With the proposed scheme, a blind sensing algorithm is introduced. The proposed scheme mainly addresses the sensing problem of secondary users as well as the primary users without any prior knowledge. Moreover, the problems coming from noise uncertainty and different kind of noise types are diminished because of added dominant noise characteristics. The performance analysis has been conducted using real data and the experimental observations have been obtained. The experimental results show that the proposed sensing algorithm outperforms the existing methods and is robust on noisy environments. By the novel method presented here, 90% true detection performance is achieved on the orders of nearly -10 dB SNR values. With nearly perfect sensing performance, the proposed CR-MAC provides simple, fast and accurate operating conditions.
Design and implementation of a teleconference system using an improved HEVC codec
Teleconferencing has become an indispensable element in any business system, because it offers the opportunity for collaborators to participate in a virtual group while remaining in divergent regions. Teleconferencing also increases productivity, minimizes travel expenses and saves travel time. This thesis presents a reliable Teleconference system that utilizes an improved high-efficiency video codec (HEVC) H.265 technology with a congestion control. The improvement is based on Coding Unit (CU) size decision and entropy coding, which provides an adequate approach to enhance the real-time video/IP technology in terms of improved video quality and increased compression ratio compared to the previous codec (H.264) and original HEVC. In the thesis, we proposed fast algorithms that reduce encoder complexity in a way more suitable for a real-time processing. Our system ensured that voice, video, text and other control packets are streamed efficiently to various individuals with smooth real-time communication. This thesis proposed a solution for a qualified server/client teleconference system. The system is an open, expandable platform and can work efficiently in an intranet/internet environment.