Department of Computer Engineering
Permanent URI for this collection
Browse
Browsing Department of Computer Engineering by Subject "computer science"
Now showing 1 - 6 of 6
Results Per Page
Sort Options
Item A COMPARISON OF IMAGE DETECTION ALGORITHMS YOLO AND FASTER R-CNN IN DIFFERENT CONDITIONS(2022-06-13) ABDULGHANI, ABDULGHANI MAWLOOD A.GHANI; Dalveren, Gonca Gökçe MenekşeIn this thesis, we compare YOLOv4 with YOLOv3 and Faster R-CNN in terms of better object detection in both challenging weather conditions and darkness. Moving objects such pedestrians, cars, buses and motorcycles can be difficult to detect in rainy, foggy and snowy weather conditions or even at night. This study is aimed at evaluating the three modules to determine which perform best in such circumstances, bearing in mind that none of them was initially intended to perform in bad weather conditions or at night. This Study is done by utilizing Tesla P4 GPU, with 12GB RAM. We trained these algorithms with an Open-Image dataset, where YOLOv4 has scored the best results at 40,000 iterations, 72 mAP, and 0.63 Recall. On the other hand, YOLOv3 has scored maximum at 36000 iterations, 65.53 mAP, and 0.54 Recall. Finally, Faster R-CNN scored 36,000 iterations, 51 mAP, and 0.49 Recall. In terms of detection performance evaluation, YOLOv4 performed at 42 FPS, while YOLOv3 was at 37 FPS and Faster R-CNN at 10 FPS in video with 30 FPS. Based on the results, YOLOv4 has performed the best in comparison to YOLOv3 and Faster R-CNN.Item ABSTRACTIVE TEXT SUMMARIZATION USING DEEP LEARNING(2022-01-11) ABBAS, HANAN WAHHAB; YILDIZ, BeytullahThe ability to produce summaries automatically helps to improve knowledge dissemination and retention, as well as efficiency in a variety of fields.There are basically two approaches to summarizing, abstractive and extractive. The abstractive approach is considered more successful as it is the process of creating a brief summary of the source text to capture the main ideas. In this approach, summaries created from the source text may contain new phrases and sentences not included in the original text. The use of attention-based Recurrent Neural Networks encoder-decoder models has been popular for a variety of language-related tasks, including summarization and machine translation. Recently, in the field of machine translation, the Transformer model has proven to be superior to the Recurrent Neural Networks-based model. In this thesis, we propose an improved encoder-decoder Transformer model for text summarization. As a baseline model, we used Long Short-Term Memory with attention, a Recurrent Neural Networks model, for the abstractive text summarization task. Evaluation of this study is performed automatically using the ROUGE score. Experimental results show that the Transformer model provides a better summary and a higher ROUGE score.Item FAST HEADER MATCHING IN NETWORK PACKETS USING FIELD PROGRAMABLE GATE ARRAYS(2022-01-17) NASER, ANWER SABAH; Özbek, Mehmet EfeThe hardware architecture of the parallel process multiple RAM that emulates the behaviors of content addressable memory for packet classification is presented in this thesis. With the increase in Internet networks’ speed, the speed of detection of intruders has become a basic requirement. In this work, a packet header field is used in a fast and efficient way to detect intruders to prevent them from accessing the data. The application test results were fast and compatible when used the FPGA board technique from Xilinx. Finally, the design, synthesis of this parallel process multiple RAM packet header detector has been achieved using Vivado 2018.2 simulator, and coding is written in Verilog HDL language and Xilinx Artix-7 FPGA (Field Programmable Gate Array) kit was used.Item New Greedy Algorithms to Optimize the Curriculum-based Course Timetabling Problem(2022-01-14) Coşar, Batuhan Mustafa; SAY, Bilge; Dökeroğlu, TanselThis thesis presents a set of new greedy algorithms for the optimization of the well-known ”Curriculum-Based Course Timetabling” (CB-CTT) problem, which is a subtype of the ”Course Timetabling” problem. The main goal of the study is to minimize the total number of soft constraint violations while preserving the satisfaction of hard constraints (feasible solutions). Since the problem is NP-Hard and large instances of the problem cannot be solved in practical times, greedy algorithms that work to produce acceptable results in a few seconds are good alternatives to brute-force and evolutionary algorithms that spend hours of execution times to search for an optimal solution. Instead of using a single heuristic as it is performed by many greedy algo rithms, we define and execute 120 greedy heuristics on the same problem instance simultaneously and report the overall best result, which would produce better results than which is obtainable by using a single greedy heuristic algorithm. The best results with respect to the No Free Lunch Theory, which states that the costs of greedy heuristics should be comparable on average, are reported. Our proposed greedy algorithms use the Largest-First, Smallest-First, Best-Fit, Average-weight first heuristics, and the Highest Unavailable course-first heuristics simultaneously while assigning the courses to the available rooms that are ordered by their capacity according to the above four different criteria. In order to evaluate the performance of our proposed algorithm, we carry out experiments on 21 problem instances from the Second International Timetabling Competition (ITC-2007) benchmark set. The experimental results verify that the proposed greedy algorithms can report zero hard constraint vio lations (feasible solutions) for 18 problems with significantly reduced soft-constraint values.Item Reinforcement Learning for Intrusion Detection(2022-01-17) Saad, Ahmed Mohamed Saad Emam; Yıldız, BeytullahNetwork-based technologies such as cloud computing, web services, and Internet of Things systems are becoming widely used due to their flexibility and preeminence. On the other hand, the exponential proliferation of network-based technologies exacerbated network security concerns. Intrusion takes an important share in the secu rity concerns surrounding network-based technologies. Developing a robust intrusion detection system is crucial to solve the intrusion problem and ensure the secure delivery of network-based technologies and services. In this thesis, a novel approach was proposed using deep reinforcement learning to detect intrusions to make network applications more secure, reliable, and efficient. As for the reinforcement learning approach, Deep Q-Learning is used alongside a custom-built Gym environment that mimics network attacks and guides the learning process. A supervised deep learning solution using a Long-Short Term Memory architecture is implemented to serve as a baseline. The NSL-KDD dataset is used to create the reinforcement learning environment and to train and evaluate the baseline model. The performance results of the proposed reinforcement learning approach show great superiority over the baseline model and the other relevant solutions from the literature.Item STUDENT ACHIEVEMENT PREDICTION BASED ON ARTIFICIAL NEURAL NETWORK VERSUS FUZZY LOGIC(2022-01-14) Al-Khafaji, Mustafa; ERYILMAZ, MeltemE-learning currently represents great importance in the process of developing the educational process in all stages from the primary classes to the postgraduate classes, as it provides an interactive graphical environment that is easy to deal with, as it attracts students to it with ease and makes them interact with it. This study, used artificial intelligence techniques, represented by both the neural network and fuzzy logic, to predict student achievement in the final exam who use the E-Learning Management System. The dataset used in this study was taken from an Iraqi engineering college, and it represents data of 200 students who have enrolled in the computer science course. The data were (gender, age, resources downloaded, videos viewed, discussion chat joined, midterm1 score, midterm2 score, final exam score). The type of artificial neural network used was pattern neural network. Levenberg-Marquardt's algorithm was used to train the neural networks. For the fuzzy logic Sugeno fuzzy inference system was used. The study results were promising and good as the results showed that the students who spend more time on the learning system have the most success rate. In this study, the neural network trained, tested, and all the results were recorded, where the accuracy of the results was 73%. The same thing for the fuzzy logic technique where the results were more accurate, as the average percentage of accuracy results was 88%.