Quick Access: Best Papers, Presented Papers
Best Papers
Hot-End Quality Control and Process Monitoring for Glass Packaging Production Using SWIR Imaging
Mehmet Can Türkeş, Nuri Erdoğan, Yiğit Berkay Uslu, Boğaziçi University
Abstract – Hot-end process monitoring and quality control systems for glass packaging industry rely mainly on thermal imaging of freshly formed, hot glass gobs emitting infrared radiation. The strength of this infrared radiation depends on various factors such as the temperature and thickness of the glass as well as the reflected radiation coming from the neighboring products. Those reflections, combined with the radiation of the glass product itself and background illumination, cause high ambient temperature and introduce significant thermal noise to the captured images, thus posing a challenge to product segmentation. In this paper, we propose a robust and real-time multi-stage segmentation method for hot-end glass product images captured by a Short-Wave Infrared (SWIR) line-scan camera under illumination and ambient temperature conditions that are not necessarily ideal. Our method combines multi-stage Otsu thresholding, which is improved by region sampling and histogram tail trimming, and marker-based watershed segmentation to obtain an initial binary mask for the product. Then, an active contour algorithm is implemented with negative distance transform of this mask as the external energy term. Resulting contours, which are observed to be robust in images even at low signal-to-noise ratios (SNR), lay the groundwork for intensity and shape-based analyses.**
Sentiment Analysis and Reporting from Text Data in Film Reviews
Egehan Eralp, Abdullah Oğuz Türk, Kerem Solmaz, Istanbul University Cerrahpaşa
Abstract – Human and communication are inseparable. For this reason, it is very important to be able to express feelings towards the other side in people and their communication. Nowadays, with the digitalizing world, people perform most of their communication in a written form in computer environment. The inability of these text data to express feelings and thoughts is a big problem in terms of communication. In order to eliminate or minimize this problem, it is very important to analyze the sentiments contained in the text data. In this context, many input representation models (Word2Vec, Doc2Vec) and text classification models (Naïve Bayes, Decision Tree, Logistic Regression, LSTM) were developed and evaluated in our study. As a result of the evaluations, the sentiment contained in the text data in the film reviews and the score corresponding to the text were estimated successfully with the most successful models. A report is presented to the user/institution along with the analysis of the user-based data generated as a result of the estimations.
Presented Papers
Hot-End Quality Control and Process Monitoring for Glass Packaging Production Using SWIR Imaging
Mehmet Can Türkeş, Nuri Erdoğan, Yiğit Berkay Uslu, Boğaziçi University
Abstract – Hot-end process monitoring and quality control systems for glass packaging industry rely mainly on thermal imaging of freshly formed, hot glass gobs emitting infrared radiation. The strength of this infrared radiation depends on various factors such as the temperature and thickness of the glass as well as the reflected radiation coming from the neighboring products. Those reflections, combined with the radiation of the glass product itself and background illumination, cause high ambient temperature and introduce significant thermal noise to the captured images, thus posing a challenge to product segmentation. In this paper, we propose a robust and real-time multi-stage segmentation method for hot-end glass product images captured by a Short-Wave Infrared (SWIR) line-scan camera under illumination and ambient temperature conditions that are not necessarily ideal. Our method combines multi-stage Otsu thresholding, which is improved by region sampling and histogram tail trimming, and marker-based watershed segmentation to obtain an initial binary mask for the product. Then, an active contour algorithm is implemented with negative distance transform of this mask as the external energy term. Resulting contours, which are observed to be robust in images even at low signal-to-noise ratios (SNR), lay the groundwork for intensity and shape-based analyses.**
Embedded System Design for Autonomous Mobile Exploration and Mapping
İlkyaz Büber, Hamdi Kılıç, Emin Yavas, Dila Özvardar, Istanbul Bilgi University
Abstract – This paper presents our work, which includes robot and image processing development to build 2D maps with the intent of moving the robot to the desired target in a static but unexplored indoor environment. The main purpose of this study is to develop a system with high precision and set up an effective pathfinding method by drawing a 2D map of the area concerned by utilizing a minimum number of low-cost components as well as avoiding the use of costly sensors such as LIDAR. Implementing the algorithms that transform high-level requests of humanity to low-level instructions of how to move is an essential need in robotics. Thus, by analyzing the various attributes and properties of an image to draw a static 2D indoor environment map, the algorithms have been developed in Spyder; herein the programming language is Python. We propose several algorithms in this work, but most significantly, an A* algorithm, that determines the shortest path possible between the robot’s initial position and desired position, thus the robot can move without hitting the obstacles by following a projected path. In this algorithm, a designed grid-map forming is used. Also, increasing the accuracy of a 2D map developed by sensor applications, field vision, and image processing. We present methods that combine the strengths of these various approaches. Consequently, through these proposed methods this work has been completed with a degree of success, thus, mobile exploration and mapping robots will be carried one step further from the old generation technologies. Hence, this work will demonstrate that robots can improve the overall mapping performance. On the other hand, robots can approach the locations to report sensor data and show more detailed views of an area where humans cannot enter directly. This study also taking into consideration the impact of these types of robots.
The Magneto-Optical Trap and Its Application in Atomic Interferometers
Ecem Nur Duman, Yeditepe University
Abstract – The work done in the field of cold atoms is very valuable for studying the basic structure of matter and merging it with current technology. One of the most widely used methods in this field is the Magneto-Optical Trap (MOT) which is a method that enables cooling the atom down to the μK range without cryogenics and trapping jointly. In this paper, both theoretical and experimental details of the MOT will be briefly explained. Laser cooling, Doppler shifting, and the Zeeman effect will be reviewed to explain the physics behind the MOT. Then Ultra-High Vacuum system and magnetic coils will be discussed as they are an integral part of the experimental setup. In addition, the application of the MOT in atomic interferometers that provides high precision and scalable technology will be given.
Comparison of Different Classifiers for Age Recognition System Based on Speech Signals
Armağan Fidan, Rabia Özge Bircan, Bahçeşehir University
Abstract – Age estimation through human speech has become one of the most popular methods in many areas such as bank security systems and IoT user interactive systems in recent years. One of the implementation types of speech analysis can be age estimation which can distinguish age groups when user identification is needed. The idea behind it is based on looking at the characteristics of speech that help determine the age of people and clarify some basic information about them. The encountered challenges in the age estimation system are identifying the features of the signals and selecting a classifier. In this study, we aim to find the best classifiers among different types of algorithms using speech features. For this purpose, eight age groups consisting of different speakers were taken without dividing by gender to test various classifiers. The methods used and the comparison results are presented in detail in the text.
Design of 4-bit and 8-bit Pseudo Noise Sequence Generators with All Zero Condition Protection Circuit
Alper Kurt, Özyeğin University
Abstract – Pseudo Noise (PN) Sequences have been utilized for modern communication and measurement systems. They can be locally generated in both transmitters and receivers. PN sequence can be generated with Linear Feedback Shift Registers (LFSR). In this paper, LFSR based 4 bit and 8 bit PN sequence generator designs are proposed. Although PN sequence generator can be implemented on FPGA with VHDL, this paper focuses on the hardware implementation of PN sequence generator with Integrated Circuits (ICs) which can be found on the market to dispose of the cost of FPGA and design a PN sequence generator block which can be used for many systems. In the hardware implementation, if the initial seed of the LFSR is in all zero condition (0 0 0 0 for 4 bit) system would be locked and the PN sequence would never be generated. To overcome this problem, all zero condition protection circuit is proposed.
Mean Energy Based Audio-Visualization Method
Halil Said Cankurtaran, IEEE Turkey PDEA
Abstract – In this study, an audio-visualization method is proposed. This method consists of five fundamental steps, (i) processing of acoustic signal,(ii) partitioning based on specified frames per second (FPS), (iii) calculation of mean energy of partitions, (iv) a primitive equalization and (v) visualization. Visualization is inspired from the Knight Industries Two Thousand (KITT) car which is a character in Knight Rider TV series
N-by-N Bit Multiplier in M Cycles
Orhun Arık, Özyeğin University
Abstract – Multiplication is one of the most important operations in embedded systems. Most operators usually take in 64bit inputs as operands. This paper focuses on N bit inputs as operands, where N will be in the hundreds if not thousands of bits. This multiplication operator will be implemented on an FPGA, which will be done in Xilinx ISE, Verilog. Due to the number of physical switches on the FPGA boards, UART connection with a PC is required for inputting the operands. Once the inputs are filled to the registers in the FPGA the multiplication operation can begin, thus the “loading” cycles of the inputs do not count towards the M cycles. The multiplication operation utilizes resource sharing for keeping the area of the circuit as minimal as possible, however, latency is not disregarded for area’s sake thus, a balance between the two will be kept. The M cycles depend on the level of resource sharing applied, which is linked with the operator architecture/algorithm. The operator architectures/algorithms that will be tested are first schoolbook algorithm then Xilinx IP Core generated multiplier, simple HDL star operator, Karatsuba and ASOHA.
Performance Evaluation of Integrated CMOS Photodiodes in 180 nm Technology for VLC Systems
Batuhan Yavaşoğlu, Cemre Kut, Istanbul Bilgi University
Abstract – The aim of this work is to compare integrated CMOS photodiode performance with the off-the shelf alternatives in a Visible Light Communication (VLC) system. We use two different photodiodes and compare them. Then through various tests, the performance of the photodiodes are evaluated to determine the most efficient component. In this work, we obtained different current and voltage values by changing the distance only by keeping the resistance the same in dark and bright environments in a BPW34 and a photodiode CMOS integrated circuit (IC).
A Vertical Federated Learning Method for Multi-Institutional Credit Scoring: MICS
Muhammed Yusuf Efe, Boğaziçi University
Abstract – As more and more companies store their customers’ data; various information of a person is distributed among numerous companies’ databases. Different industrial sectors carry distinct features about the same customers. Also, different companies within the same industrial sector carry similar kinds of data about the customers with different data representations. Cooperation between companies from different industrial sectors, called vertical cooperation, and between the companies within the same sector, called horizontal cooperation, can lead to more accurate machine learning models and better estimations in tasks such as credit scoring. However, data privacy regulations and compatibility issues for different data representations are huge obstacles to cooperative model training. By proposing the training framework MICS and experimentation on several numerical data sets, we showed that companies would have an incentive to cooperate with other companies from their sector and with other industrial sectors to jointly train more robust and accurate global models without explicitly sharing their customers’ private data.
Evaluation of Behavior Tree and Finite State Machine based Artificial Intelligence Algorithms in Shooter Games
Özgür Çebi, Yeditepe University
Abstract – Nowadays, artificial intelligence algorithms have a major role in our lives. One of these areas are computer games. Most of the games has AI-controlled components in order to control the game behaviour. It is a must to keep these algorithms up to date in order to provide the best experience for the users. This project will focus on the importance of artificial intelligence algorithms in shooter games. The game will consist of two teams with the same number of agents. Each team will use a different algorithm to control the agent behaviour which are Finite State Machine (FSM) and Behaviour Tree (BT) algorithms. FSM algorithm will contain the core mechanics of the game which will be simplified to eliminating the enemy team. BT algorithm on the other hand will evaluate the environmental factors and make tactical decisions which is expected to increase the win rate. Results will be used to determine the best fitting algorithm for shooter games. Two main mechanisms will be included in both algorithms which are movement decision system and enemy targeting system. To test the algorithms, a shooter game will be created with Unity to gather the necessary data. The agents will be bound with the same rules such as morale system and weapon accuracy system. There will also be environmental factors such as covers which reduce the probability of being hit. The game will be over when a team has lost all of its agents. Algorithms will be tested on different maps. Win/lose ratio, game time, remaining number of agents and map design will be used to evaluate the results.
Handwritten Text Segmentation and Optical Character Recognition with Transformer Architecture
Onur Kirman, Özyeğin University
Abstract – Despite the technological revolution and printing press, most of the documents mankind has generated are still in handwritten format. Precise segmentation of free-form and curvy text is utterly important since it is the phase that defines the final quality of the OCR. We present a new pixel-wise handwritten text line segmentation method. We benchmark rigorously with the recent de-facto architectures. We present solutions employing Attention via Transformers by using the IAM Handwriting Database. After the dataset is prepared for training, validation, testing; we first demonstrate the performance of the widely used and accepted model U-Net. Then, we discuss possible improvements over the convolutional model and propose two novel architectures, TransUNet and Attention U-Net that both perform around 1% better than an extensively fine-tuned U-Net model. We visually compare our models’ line separation quality. They visually outperform U-Net and separate the text lines flawlessly even at the same level of pixel-wise accuracy of the U-Net model. Finally, we show a comparison with the Transformer based DETR method.
One-shot Learning for Retail Product Recognition
Orhan Eren Akgün, Boğaziçi University
Abstract – There is a lot of interest in the recognition of retail products based on their images and automatic processing of this information. Currently, deep learning motivated algorithms such as deep convolutional neural networks (CNN) are the state-of-the-art. However, training deep neural networks notoriously requires a large amount of data. A large dataset is not always available for retail products since new products are frequently introduced and data collection and annotation is an expensive task. Moreover, available data during training might be packshot images or digitally crafted visuals of the products, which introduces a domain shift during test time. In this work, we compare the performance of three different models, a CNN with softmax-cross entropy loss, a CNN with cosine similarity based loss, and a variational prototyping-encoder (VPE), on one-shot product classification problem. We introduce and evaluate our approach on a new retail product dataset, which consists of 37 classes and a total of 14,668 automatically cropped product image samples obtained from actual supermarket shelf scenes. The models are evaluated using one-shot training examples from both in-store and digital product images to demonstrate the effect of domain shift on performance. The dataset we introduce can provide a more realistic estimate of the model performances since it includes images of very similar objects that are subject to partial occlusion, pose and illumination changes.
Sentiment Analysis and Reporting from Text Data in Film Reviews
Egehan Eralp, Abdullah Oğuz Türk, Kerem Solmaz, Istanbul University Cerrahpaşa
Abstract – Human and communication are inseparable. For this reason, it is very important to be able to express feelings towards the other side in people and their communication. Nowadays, with the digitalizing world, people perform most of their communication in a written form in computer environment. The inability of these text data to express feelings and thoughts is a big problem in terms of communication. In order to eliminate or minimize this problem, it is very important to analyze the sentiments contained in the text data. In this context, many input representation models (Word2Vec, Doc2Vec) and text classification models (Naïve Bayes, Decision Tree, Logistic Regression, LSTM) were developed and evaluated in our study. As a result of the evaluations, the sentiment contained in the text data in the film reviews and the score corresponding to the text were estimated successfully with the most successful models. A report is presented to the user/institution along with the analysis of the user-based data generated as a result of the estimations.
Aerial Object Tracking with Centers
Hakkı Motorcu, Özyeğin University
Abstract – Multiple Object Tracking Task has become more popular due to rise of Deep Neural Networks and increased computation power. However Aerial Object Tracking task remains still challenging due to model scalability, model speed, high resolution inputs, and relatively small targets for tracking. Normal object trackers use anchor box-based detectors and detection matching algorithms to track objects. This approach introduces complexity due to repetitive anchor box detection method. To reduce complexity, we used recently emerged center-based object detection method and next frame object position prediction by using previous images and interior feedback loops. Our one stage model does object tracking operations without compromising accuracy while running close to real time video speed (20-21 FPS in HD) which varies with the input footage resolution.
Analysis of Turkish Sign Language for Computer-Based Phonetic Transcription from Turkish-to-Turkish Sign Language
Sueda Gülgün, Ceyda Manav, Istanbul Commerce University
Abstract – Natural languages are generally divided into spoken languages such as Turkish, English, Chinese, and sign languages such as American Sign Language (ASL), British Sign Language (BSL), Turkish Sign Language (TID). The translation of natural languages to each other can be achieved with computer-based systems. These systems translate from the source language to the target language within the framework of rules. For the accuracy of the translation, the rules of the language should be determined within strict boundaries. The purpose of this study is to analyze the distinctive grammatical structure and appearance of Turkish and Turkish Sign Language (TID), which are natural languages.
IPFS Pinning Service
Kaan Topçu, Yeditepe University
Abstract – InterPlanetary File System (IPFS) is a distributed file system which allows it’s participants to share, host and store data in a distributed way. In IPFS there is a key content called pinning. Unpinned files are removed from the system by the scheduled garbage collection. Only way to keep data available all the time is to pin them. It can be achieved in two ways: either using local storage or a 3rd party service. This project aims to build an IPFS pinning service over IPFS network. Project has its own database to control user management, pinned datas’ informations and encrypted hashes. Pinning operations can be done and can be visualised via simple web UI. In order to simulate the system in local, client-server model is used. The service is implemented as a pilot system and evaluated.
OpenCache: An Open-Source Cache Generator
Eren Doğan, Özyeğin University
Abstract – Designing hardware manually is difficult for researchers since it is time consuming and difficult to scale these hardware designs. Therefore, the interest on open-source hardware has increased in the last decade. Recently, Guthaus et al. published their project of OpenRAM, which is an open-source SRAM compiler, for researchers to design SRAM arrays easily. Similarly, the aim of OpenCache is to provide an open-source tool for researchers who need custom cache for their hardware design. OpenCache generates behavioral Verilog modules for cache logic and configuration files for OpenRAM to generate internal SRAM arrays of the cache.
