2019-2020

2019 - 2020 Spring

Abstract: Agriculture is vital not only because it provides food for the domestic market but also is a strong means of international trade for Turkey where 26.82% of all land is arable. Storage and transportation of harvested goods should be handled carefully in order to prevent spoiling. The maintenance of the goods is done by cooling, humidity control etc. All these actions require power and transmitting power to the entire farming field is expensive. The system works as an off-grid system and uses solar energy to produce its own electricity by PV panels. The system uses a temperature control algorithm to keep the goods inside fresh while reducing power consumption. This system is remotely accessible by users. Users can check the temperature of the storage. Users can also tell the system what kind of good they want to put in the storage so that the system can arrange the temperature of the storage for the selected good. Communication between the user and the system is done by a mobile app. The system is composed of 200W panels, and the battery of 12V. Overall power consumption for one day is 900W.

Abstract: Nowadays, mains electricity alarm systems are used very often and with the technological developments, alarm systems work in a comprehensive manner. Even, the thief can be photographed with the included imaging system in these alarm systems. However, thieves have become aware of this evolving technology and have taken precautions. Before entering the houses and workplaces, they turn off the electricity and disable the alarm system. In this way, they steal the valuable goods in the houses and workplaces without any alarm system working. We build a system that works with renewable energy in order to solve this situation and to create a low-cost system. With the use of panels in the homes to obtain energy, no tools are connected to the mains electricity. In this way, even if thieves cut the electricity and break into any workplace or house, the alarm system will be always active as it operates with the sun which provides unlimited energy. Thus, when people are away from their homes, summer houses and workplaces, they will be able to continue their lives comfortably. In this thesis, the design and prototype of the alarm system that works with solar energy is made by using Raspberry Pi 3B+. Passive infrared (PIR) sensor has been activated and tested. The Third Generation (3G) modem, which meets the internet needs of the system, has been adapted to the system. Solar energy is used to provide the energy of this alarm system, and Raspberry Pi Camera is used for recording video. This video is 10-second. After this video is recorded, it will send to user’s email.

Abstract: The main objective of this project is to develop automatic speech recognition (ASR) and highlighting systems for lecture videos. ASR is the method of deriving the transcription (word sequence) of an utterance, given the speech waveform. The purpose of highlighting system is take attention to important information in a text by generating higlighted sentences. Nowadays there are lots of open online courses accessible to everyone in digital platforms. The use of videos for educational purposes provides convenient and flexible learning opportunities to learners both inside and outside the classroom. Although watching videos is one of the most effective ways of learning, the major disadvantage of the video-based education systems is that interaction between learner and video is limited. It reduces the maximum efficiency of online videos. Since ASR and highlighting systems can increase the interaction by returning highlighted transcriptions of video clips, developing these systems is quite important to facilitate learning from online course videos and ASR makes these resources available more efficiently. In the project, the ASR system was developed for lecture videos at MEF University. The system was implemented with Kaldi which is a publicly available ASR toolkit. Then, transcriptions were used to provide highlighted sentences. The summarization concept was used to create highlighted sentences in this project. To provide highlighted sentences, the Bidirectional Encoder Representations from Transformers (BERT) was used as a continous space word vector representation and the K-means algorithm was used in order to detect important sentences in the text. In addition to the textual feature, the acoustic features of the speech signals was used for developing highlighting system. Then, a rule-based decision-making system that use information comes from k-Means and acoustic importance values, was developed for sentence selecting process to use in extractive summarizer for highlighting system. As a result of the project, the developed system is applied to the lecture videos and this result is encouraging in terms of automatic highlighting of speech recognition outputs where sentence boundaries are not defined explicitly. Textual representation yields promising result but no further improvements were obtained with acoustic features. The reason for the unexpected result is that the speaker speaks in lecture videos monotonically. This causes the pitch values which we used as an acoustic feature for each sentence, were not valuable to increase the performance of the highlighting system.

Abstract: In recent years, in the area of machine learning, the methods used when training the neural network models have increased and expanded. In this project, we have developed a laboratory report grading system that will give instant feedback to students. To do this, we need to apply a neural network model that can automatically grade some comment questions in laboratory reports. Firstly, the answers in the comment questions were translated into machine language so that they could be used. Using the word embedding technique, we have made the vector representations of the answers proper for the neural network model. Then we trained the recurrent neural network model with these representations to solve the text-regression problem. We also designed a user-friendly interface so that students can use it.

Abstract: Smart home systems integrate many new technologies to improve people's quality of life through the home network. In this project, a smart home assistant whose name is Jarvis is developed like Google Home or Amazon’ Alexa. The assistant presented for people with disabilities. This smart home assistant understands previously defined Turkish commands, and then it performs the tasks associated with the given commands. The system has two main parts which are Automatic Speech Recognition (ASR) and performing tasks related with commands. The Mel-Frequency Cepstrum Coefficient (MFCC) and Dynamic Time Warping (DTW) methods were used for the ASR part. In the second part two circuits were designed. One of the circuits works with 2 different switches that are set up to turn the lamp on and off because when the user turns off or on the lamp the user can on or off it from the normal switch. The second circuit has an IR Receiver to decode signals of Tv remote controller and an IR Transmitter to send this signal to Tv by speech command. A package for hotword detection was used. The system is always listening to users and when users say ‘Jarvis’ command, it gets active and waits for next commands. The accuracy of the system was tested. The results showed that accuracy of performing commands of design is more than 80%.

Abstract: This paper describes a secure-pass system with signature verification. The secure-pass system works on a field-programmable gate array (FPGA) with the help of a touchpad. The main purpose of this project is to improve the secureness of the secure-pass systems. An online signature verification method that is based on extracted features of a signature is used in this project. The implementation of the system consists 4 main stages; data acquisition where data is collected by touchpad and pre-processed, feature extraction where gathered data is transformed into set of different features, dynamic time warping (DTW) for similarity measurement between signatures and decision for authentication based on the DTW measurement results. The algorithm was tested using database of 4 users and 200 signatures, and verification rates obtained for each users’ genuine and forged signatures. 86% accuracy was achieved after the tests.

Abstract: Recent developments bring the spiking neural networks as the third version of the artificial neural networks. This new neural network structure has many benefits in comparison with the second generation of neural networks. Benefits can be listed as power consumption on neuromorphic hardware, more reliable on continuous data and powerful computation in the theory. Spiking Neural Networks are ongoing research area. Our aim in this field, developing a Spiking Neural Network on FPGA by using the advantages of spiking neurons based on their similarity to biological neurons and running this network on a simple pattern recognition application. A parametric spiking neuron model is implemented on FPGA (Cyclone V) and network with ten neurons implemented. Four different patterns are classified by spike counts of the network and it was observed that the whole network occupies 55% area on FPGA.

2019 - 2020 Fall