A MICROWAVE TOMOGRAPHY ALGORITHM FOR MEDICAL IMAGING
Advisor: Asst. Prof. Egemen Bilgin
Team: Didem Ertek, Gökhan Küçük, Yağız Ertan Gürelli
Keywords: Microwave medical imaging, microwave sensing, microwave tomography, differential imaging
Abstract: Most medical imaging systems used today depend on ionizing radiation such as magnetic resonance (MR), ultrasound, X-rays, and computed tomography (CT). A microwave imaging system can be an alternative to other medical imaging systems. It can be used in medical applications for the diagnosis of disease within the human body as an alternative or complementary imaging system. Microwave imaging offers many desirable features as a tool for cancer assessment. Microwave imaging is provided by non-ionizing electromagnetic signals and lower cost than current methods. These advantages could result in safer and healthier testing. Also, designs can be made portable for microwave medical imaging, making real-time results viable for applications such as stroke detections. An algorithm that can gather real-time results completely safely with non-ionizing signals and a much lower implementation cost will be developed in this project. This algorithm will be simulated on MATLAB software with realistic phantoms.
ACOUSTIC WAVE INVERSE SCATERRING
Advisor: Asst. Prof. Egemen Bilgin
Team: Başar Utkan Sarıtaş, Hayati Mert Eğribel, Yasin Mert Erkekli
Keywords: Scattering, Acoustic Wave
Abstract: Acoustic waves are waves that can propagate in solid, liquid and gas with a wave pattern. Acoustic waves have wavelength, frequency, amplitude and period. The propagation and vibration directions of acoustic waves are the same. The scattering problem occurs as a result of the interaction of electromagnetic fields with matter. In a homogeneous medium, scattering does not occur for a wave that coincides with an empty space. For an acoustic wave to scatter, it must interact with an object in an electromagnetic field. scattering is the reflection of an object from its rough surface, with scattering a certain area is scanned. Since acoustic waves can spread in air, water and gas environments, they can be used in many areas, for example, cancer cell detection for medicine or as sonar in submarine vehicles in the military field. With direct acoustic wave scattering, the frequency to be used in the imaging process has been determined and the imaging process has been performed with the values that agree with the direct acoustic wave scattering in medical imaging with reverse acoustic wave scattering. Different methods have been used for imaging with reverse acoustic wave scattering. RTM acoustic, TSVD, LSM and music methods are used for imaging.By experiencing these methods in the simulation environment, it was seen which method works most efficiently.A clear result was seen in 4 of the 5 methods we used while working on acoustic inverse scattering.Except for the TSVD method, other methods were able to detect what the cell was inside. It was clearly observed which type of cells (cancerous, tumorous) were present in the cell.
BURIED OBJECT DETECTION
Advisor: Asst. Prof. Egemen Bilgin
Team: Alp Kutay Atasever, Ibrahim Berhak Onder
Keywords: Buried Object, Signal Detection, Scattering, Electromagnetic, Electric Field, Inverse Scattering
Abstract: The signal processing strategy that employs field models for the air, soil, and object environments has the highest chance of addressing the buried item detection problem. As a result of these limits, standard threshold detection algorithms may not be able to tackle the hidden object identification problem satisfactorily, especially when the item is small and the return signal is weak. These assumptions allow us to use accurate electromagnetic field scattering theory to address the concealed object detection problem and design a computationally efficient optical detection system. The difficulty of detecting known buried artifacts and estimating their location using electromagnetic field data is crucial in many technological domains, such as demining, buried waste remediation, excavation planning, and archaeological investigations. Our aim is to identify the electric field formed by scattered particles as they scatter, and we'll use mathematics to calculate the field created by items in the bottom half- space of the measuring region above. The presence of a large number of randomly scattered unwanted things, whose returns obscure the return from the object of interest, as well as random roughness of the air/soil interface, results in incoherent returns. In particular, while describing the object, the Scattered Field will be determined first and then, the Inverse Scattering problem will be solved to depict the buried objects.
DRIVER DROWSINESS DETECTION AND WARNING SYSTEM
Advisor: Asst. Prof. Tuba Ayhan
Team: Arda Özdemir, Berkay Yazıcı
Keywords: Artificial Intelligence, Machine Learning, SoC, Driver Drowsiness Detection
Abstract: The aim of this project is to detect the drowsiness level of the driver in the vehicle, to warn the driver and to prevent possible accidents. Facial Landmarks, Eye Aspect Ratio (EAR) and Haar Cascade methods were used to conduct the project. The system firstly tested on a computer. Then it was implemented using the PYNQ-Z2 development board. High accuracy and response time achieved as targeted. In addition, the audio warning system and the function of sending the instant status of the driver to the cloud system were successfully implemented. In this way, a system that can work in the vehicle and prevent the driver from having an accident is designed.
PROSTHETIC HAND CONTROLLER
Advisor: Asst. Prof. Tuba Ayhan and Asst. Prof. Yusuf Aydın
Team: Berkay Kürkçü, Eren Naçar, Onur Berk Sunal
Keywords: EMG Sensor, Machine Learning, Prosthetic Hand, Dataset, Prototype, Development Board Kit
Abstract: An inexpensive prosthetic hand controller design was considered to facilitate the daily life routines of amputated individuals. The scope of our project is to what extent it is possible to reduce the number of conventional sensors used in prosthetic hands and how the system can be made more accessible in terms of price. Within the scope of this project, EMG sensors were used to detect muscle activations and hand movements. The type of EMG sensor that has been chosen is the Surface EMG sensor. This sensor mainly consists of 2 main units. These are electrodes and signal conditioning circuit. According to specified movements parameters, data set for the machine learning algorithm will be created to identify gestures. The motor positions will be controlled by the motor controller. Prototype of the project has been also realized by Nucleo64 development board kit.
HOTWORD DETECTION
Advisor: Asst. Prof. Ebru Arısoy Saraçlar
Team: Emirkural Koçer, Hasan Duyan
Keywords: Hotword, Detection, Deep Neural Network, Mel Frequency Cepstrum
Abstract: Hotword detection is an algorithm that monitors an audio stream for a hotword and activates your voice assistant when it detects it. It allows detecting the keyword in spoken words and activating the voice interface. Hotword Detection system should listen to spoken audio continuously without interruption, ignoring non-keyword words, but still triggering accurately and instantly on the keyword. In this project, our goal is to identify the predetermined word or set of words with as much accuracy and consistency as possible. The biggest reason and advantage of using the hotword detection; It works with high accuracy regardless of multiple conditions, without connecting to Wi-Fi technology or internet. The hotword detection system uses deep neural networks to achieve high performance. In the past, it used the Hidden Markov Model technique to realize the special word detection technology. Today, with the development of technology, Recurrent Neural Network technique is used. In this project, we used the fundamentals of the Digital Signal Processing course. In the sound processing part, we used various libraries such as librosa. We will also implement this project using the Deep Neural Network and Keras platform. We extracted the MFCCs of the words we used and trained them with the deep neural network with the parameters we determined with the MFCCs we extracted. The frame level accuracy of the model we trained is 94.44%. We achieved an accuracy of about 60% in the tests we performed in the decision-making tests. We can see in which sequence of the audio files we concatenate keywords and in which sequence other words appear. As a result, at the end of the semester, we got the sound data to be received and processed in Python, the feature extraction part, which is the first part of the block diagram in our project, the part of training our model, the decision making with the model we trained, and finally, after the necessary coding, whether the keyword is passed in the sound file.
HUMAN DETECTION & COUNTING
Advisor: Asst. Prof. Serap Kırbız Şimşek
Team: Sait Talha Ulutaş, Fırat Koç
Keywords: HOG, Human Detection, Counting
Abstract: In this project, we present a real-time system for detecting, counting, and tracking people in public areas. While using this project, we started as a software project, minimize the cost, and to do it with technological knowledge. As a result, we made a camera capable of human detection & counting, which can be shot on the camera module and observed on the computer screen.
It is a project we prepared by using the information we learned from electrical and electronics engineering courses and applying the Covid-19 measures according to the rules. In our study using the Histogram of Oriented Gradient method (HOG), we created a system that frames people in the green square on the camera and gives an instant alert on the screen when there are too many people. It is a project that can be changed according to the needs of the society and the wishes of the person and is developed with the knowledge of the code.
MULTICOPTER DRONE DESIGN
Advisor: Asst. Prof. Yusuf Aydın
Team: Barış Durak, Burak Çevik, Umutcan Argun
Keywords: Hexacopter, Drone, Cargo, Transport, 3D Printer
Abstract: Drones are unmanned aerial vehicles which are guided by either remote control or autonomous flight options. Although they were used just for military purposes following their first emerge, drones are used today in many fields of work including photography, filming, advertisement, transportation and the sectors of agriculture and energy. In that project, the cargo shipment with drones will be inquired; also its aim is to achieve lightweight cargo shipment with drones between predetermined stations in a city by a safe, fast and inexpensive way. Such a necessity of performing the cargo shipment with drones has further emerged with the burst of the Coronavirus outbreak and biggest shipping companies such as Amazon and UPS have started to build drone fleets to perform cargo delivery tasks. In that project, our aim is to expand drone cargo shipment further such making drone cargo shipment available in every country, particularly Turkey. Delivering a shipment into a specific address is not much safe and feasible due to drones would lack a safe landing zone such like a garden of a property, many addresses could fall within no possible appearance of a safe landing zone such like multi-storey apartments and apartment complexes in densely populated urban areas, in a such situation, the option of drones could deliver their cargo into a predetermined safe landing stations that we aim to build, being nearest to the delivery address could be considered; placing landing stations can also cease the necessity of cargo being delivered in hand at working hours.
SEALED LEAD-ACID BATTERY CHARGER WITH STATE OF CHARGE ESTIMATION
Advisor: Asst. Prof. Tuba Ayhan
Team: Hasan Can Bayrak, Anıl Dilber
Keywords: State of Charge, Sealed Lead-Acid Battery, Battery Charger, Hall Effect
Abstract: The importance of rechargeable batteries is increasing day by day. Based on this, in our project, we designed a system that displays the charge value of our battery with the voltage, current and state of charge method by ensuring the charge of the lead acid battery.
SMART DOOR LOCK SYSTEM WITH FACE RECOGNITION
Advisor: Asst. Prof. Serap Kırbız Şimsek
Team: Cansu Yılmaz, Çağla Su Taşdemir
Keywords: Biometrics, Face Recognition, Local Binary Patterns, Eigenfaces, Feature Extraction, Face Detection, Viola-Jones Algorithm
Abstract: Biometrics refers to the physical and behavioral features that distinguish people from one another. Face, fingerprint, iris, voice, and DNA are examples of physical biometrics. Fingerprint identification, face & voice recognition technologies are built using these attributes and are used to identify individuals in a variety of settings such as security systems and smartphones. Along with the developing technology, face recognition technologies are used in many systems in order to both facilitate daily life and increase security. A human face in a digital image or video frame is compared to a database of faces using face recognition technology. In this project, it is aimed to adapt face recognition technology to door lock systems and develop a solution that is both safer and more practical compared to the use of traditional keys. Problems such as not being able to enter the house due to forgetting or losing the key, security concerns caused by the key being stolen by strangers, etc. are expected to be eliminated. In this system, faces are used as keys. In the project, a face recognition system is developed using MATLAB. The system uses Viola-Jones Algorithm (Eigenfaces) for face detection and the Local Binary Patterns (LBPs) Method for feature extraction. The system connects to the camera and gets the current frame from the real-time video every 0.5 seconds. The system compares the face in the input image to the ones in the database. If the input image matches one of the images in the database, the servo motor rotates the door lock 90 degrees and the door becomes unlocked. The lock stays in the open position for 5 seconds, then the servo returns to the 0 degree position and locks the door. If the input image does not match any of the faces in the database, the door remains locked.
SOLAR POWER MONITORING SYSTEM
Advisor: Asst. Prof. Tuba Ayhan, Asst. Prof. Arif Mustafazade
Team: Zeliha Turan, Ani Takıcı
Keywords: Solar Power, Solar Panel, Power Monitoring
Abstract: The cost of renewable energy systems is going down day by day, which encourages governments and large companies to invest in renewable energy systems, mostly in solar energy. As a result, great scale solar power plants are installed and employed all around the world. Generally, these plants are not placed in easily accessible locations. In fact, even the solar panels used in our homes are not easily reachable since they are placed on top of the buildings. This generates a maintenance problem.
This project aims to solve the problem by constantly measuring the generated power by the solar panel. This way, an unusual measurement will be noticed, and necessary action will be taken. Without being in need of checking the solar panel every once in a while, to see if it is working properly, the panel will be checked only when needed. This procedure will reduce the cost of maintenance and will be employed as a fault detection system.
Another aspect of this project is that the constantly measured data will be displayed on a website. This part of the project aims to raise awareness in renewable energy systems.
SPEAKER DIARIZATION
Advisor: Asst. Prof. Ebru Arısoy Saraçlar
Team: Burak Atal, Yusuf Can Fidan
Keywords: Voice Activity Detection, Diarization, Energy-Based Voice Activity Detection, Neural Network, Short-Time Energy, Deep Learning, LSTM, D-vector, Clustering, K-Means
Abstract: Speaker diarization is a system of recognizing the identity of the speaker using human speech with the help of a program. Detects speaker identity by detecting speaker changes during audio streaming. It is used to find the answer to the question "Who spoke when?" Today, it is used in many areas such as meetings, lecture attendance, and call center. Speaker diarization consists of two parts: voice activity detection (VAD) and diarization (separating sounds homogeneously). The project was completed by diarization behind voice activity detection first. This report describes the voice activity detection algorithm and the homogeneous separation of sounds. There are many methods of VAD, and an energy-based voice activity detection system have been used while doing this project and a high success rate is obtained. A python software application was used to make an energy-based VAD system. Using free and publicly available data, the energy levels of the sounds in the recording are calculated and compared with the threshold value, and the speech and silence parts of the signal are determined. If the energy is higher than the threshold value, it is erceived as speaking, if the energy is below the threshold value, it is perceived as silence. In the diarization part, parts of speech are taken and the features are extracted. After the features extraction, the LSTM model, which is the Deep Neural Networks model, is trained. Clustering is applied by taking dvectors from the last hidden layer of the LSTM model. D-Vectors have the specific properties of each data. Clustering creates output by classifying the data according to its properties. In this project, speakers were grouped using K-means clustering. Thus, it is determined where the speaker changed, that is, "who spoke when".