Saed Alqaraleh and Hatice Meltem Nergiz Sirin. MULTIMODAL CLASSIFIER FOR DISASTER RESPONSE
Abstract: Social media data has a massive difference in making correct decisions, especially in time-critical situations and natural disasters. Social media content consists of messages, images, and videos. In some cases, understanding the damage caused by disasters from text only is not enough. The exact situation and the effect of disaster are better understood using visual data. Researchers widely use text datasets, and a limited number of studies have focused on using other content, such as images. This is due to the lack of multimodal datasets and the limited number of annotated image datasets related to disasters. Therefore, we addressed this limitation by collecting disaster-related Turkish texts and their related images from Twitter. Then, using three evaluators and the majority voting, each sample was annotated as disaster or not disaster. Next, multimodal classification studies were carried out with the late fusion technique, where the BERT embedding approach and a pre-trained LSTM model are used for classifying the text, and a pre-trained CNN model is used for the visual content(images). Overall, concatenating both inputs in a multi-modal learning architecture achieved an accuracy of 91.87% compared to early fusion, which achieved 86.72%.
Renjith V. Ravi, S. B. Goyal and Chawki Djeddi. Image Encryption using Spined Bit plane Diffusion and Chaotic Permutation for Color Image Security
Abstract: With the tremendous rise in multimedia output over the last ten years, image encryption has grown in importance as a component of information security. It is challenging to secure images using conventional encryption methods be-cause of the intrinsic image properties that distinguish image cryptography from text cryptography. This study suggests employing binary key images for lossless encryption of color images. First, chaotic sequences from the 2D Henon map will be used to permute the locations of the pixels in the plaintext image. The key images, in this case, are bit planes created from an additional image. In order to improve the encryption image and make crack-ing more challenging, the bit plane will be further rotated. It concentrates on two methods for effective image encryption: bit plane slicing and bit plane spinning. In terms of a number of statistical tests, such as key experiments, information entropy experiments, and encryption quality testing, the simula-tion analysis also shows that the suggested technique is lossless, safe, and ef-fective. Based on the findings, it can be shown that the new method takes less time to conceal the remaining intelligence than it does to encrypt the complete image.
Önder Polat and Sema Koç Kayhan. Hardware Implementation of MRO-ELM for Online Sequential Learning on FPGA
Abstract: This paper presents a parallel hardware accelerator for an online variant of the extreme learning machine (ELM) algorithm, called mixed-norm regularized online ELM (MRO-ELM). ELM is a training algorithm for feedforward neural networks that has been widely adopted in the literature. The proposed parallel architecture is implemented on a field-programmable gate array (FPGA) and designed for classification tasks. It is designed to be scalable and reconfigurable for different problem sizes, and can be used for various numbers of hidden neurons. Among the existing studies, the proposed architecture is the first hardware implementation that has norm regularization with parallel processing capability. The implementation results for the proposed hardware accelerator are reported in terms of hardware efficiency and they show that the proposed design has lower resource utilization than existing parallel implementations.
Lial Alzabin, M.Jannathl Firdouse and Baidaa Khudayer. A Literature Survey on Event Detection for Indoor Environment using Wireless Sensor Network
Abstract: Event detection using a wireless sensor network plays a vital role in environmental accidents such as fires, gas leaks, and explosions occurring at random and unexpected times. Hence, one of the most critical duties of a system is to determine whether a given event will occur or not. To accomplish this, the system must be able to collect and infer environmental data that will lead to isolating and recognizing certain events. Many alternative event detection approaches are used in wireless sensor networks (depending on how the environmental data is obtained) such as that can be handled by individual sensors, isolated groups, or fusion centers. This paper provided an overview of event detection mechanisms and challenges as well as the hypothesis and covered the fundamental requirements for sensing systems. Furthermore, relevant research work on environmental event detection in current event detection approaches is reviewed and discussed. Thus, the aim of this paper is to conduct a comparative analysis of the event detection mechanism using fusion center-based fuzzy logic for improving the detection system's performance.
Ahmed Mouthanna, Sadeq Bakhy and Muhannad Al-Waily. A Computer Presentation of the Analytical and Numerical Study of Nonlinear Vibration Response for Porous Functionally Graded Cylindrical Panel
Abstract: The current study uses a new analytical model and numerical method to present a free vibration investigation of a simply supported, functionally graded cylindrical shell panel. The FG thickness properties are supposed to be porosity-dependent and vary in the thickness direction based on power-law distribution. This work makes a contribution by analyzing the performance of porous FGMs, which are employed in a particularly wide variety of biomedical applications. The governing equations are based on the first-order shear deformation theory to find the free vibration characteristics and the nonlinear vibration response by utilizing the Galerkin technique with the fourth-order Runge Kutta approach of an imperfect FGM cylindrical shell panel and include different parameters. Parameters included are the power-law index, graded distributions of porosity, and FG thickness. A numerical examination employing the finite element method and the modal investigation was conducted with the assistance of the ANSYS 2021-R1 software to validate the analytical technique.
Gizem Özgül, Şeyma Derdiyok and Fatma Patlar Akbulut. Turkish sign language recognition using a fine-tuned pretrained model
Abstract: Many members of society rely on sign language because it provides them with an alternative means of communication. Hand shape, motion profile, and the relative positioning of the hand, face, and other body components all contribute to the uniqueness of each sign throughout sign languages. Therefore, the field of computer vision dedicated to the study of visual sign language identification is a particularly challenging one. In recent years, many models have been suggested by various researchers, with deep learning approaches greatly improving upon them. In this study, we employ a fine-tuned CNN that has been presented for sign language recognition based on visual input, and it was trained using a dataset that included 2062 images. When it comes to sign language recognition, it might be difficult to achieve the levels of high accuracy that are sought when using systems that are based on machine learning. This is due to the fact that there are not enough datasets that have been annotated. Therefore, the goal of the study is to improve the performance of the model by transferring knowledge. In the dataset that was utilized for the research, there are images of 10 different numbers ranging from 0 to 9, and as a result of the testing, the sign was detected with a level of accuracy that was equal to 98% using the VGG16 pre-trained model.
Hassan Imani, Md Imran Hosen, Vahit Feryad and Ali Akyol. Efficient Object Detection Model for Edge Devices
Abstract: Deep learning-based object detection methods demonstrated promising results. In reality, most methods suffer while running on edge devices due to their extensive network architecture and low inference speed. Additionally, there is a lack of industrial scenarios in the existing person, helmet, and head detection datasets. This research presents an efficient tiny network (ETN) for object detection that can perform on edge devices with high inference speed. We take the YOLOv5s model as our base model. We compress the YOLOv5s object detection model and minimize the computation redundancy, and propose two lightweight C3 modules (MC3 and SC3). Additionally, we construct two novel datasets: H2 (consists of safety helmet and head) and Person104K (consists of person) that fill the gaps in the earlier datasets with various industrial scenarios. We implemented and tested our method on Person104K and H2 datasets and achieved about 50.6 % higher inference speed than the original YOLOv5s without compromising the accuracy. On the Nvidia Jetson AGX edge device, ETN achieves 42 % higher FPS compared to the original YOLOv5s.
Varun Deshpande, Vigneshwaran Pandi and Vishwak Nama. Smart Locking System using AR and IoT
Abstract: Modern smart door locks are more prone to damage and attacks, which will reduce robustness. Most smart door locks rely on passcode entry or face recognition outside the door which makes it easy to track and more vulnerable to attacks. Extended (XR) is an emerging term used to denote the amalgamation of multiple immersive technologies such as augmented reality (AR), virtual reality (VR), and mixed reality (MR). Recent research revealed that more than 60% of respondents believed XR will be mainstream in the next five years. Considering the features provided by the XR in terms of visualization and human interaction, XR has a wide range of applications. In this article, we propose the integration of XR with IoT devices to create a Smart door locking system that will be operated using smart glasses or mobile devices. The proposed system aims to implement a smart door-locking system to overcome physical attacks, as no part of the lock is physically intractable. MR is used to take the secret code as input from the user through Smart glasses or mobile devices, which is verified by the mobile application and the corresponding control signal to unlock the door is sent to the NodeMCU if the password is correct. The contactless feature of the lock makes it suitable in hospitals to prevent the spread of diseases and prevent users from touching radioactive components in radioactive areas.
John Jiron, Robin Gerardo Alvarez Rueda, David Vega, Felipe Grijalva, Pablo Lupera and Antonio Flores. Simulation of a Wheelchair Control System based on Computer Vision through Head Movements for Quadriplegic People
Abstract: People with quadriplegia rely on someone else to get around using a manual wheelchair. Using motorized wheelchairs gives them some independence and, simultaneously, the need of methods to control them. In the state-of-the-art, there is a great variety of these minimally invasive methods that use external devices, for example, accelerometers, which could generate discomfort. The present work proposes a non-invasive system based on artificial vision is proposed that does not require placing any device on the user. The proposed system consists of an image acquisition stage, one for face detection using Viola-Jones algorithm and another for tracking using Kanade-Lucas-Tomasi (KLT) algorithm. Additionally, a way to enter or exit the commands mode is proposed so that the user can activate/deactivate the system as required. For this, a specific movement is presented, as long as this movement is not performed, the user can move his head freely without activating the motors in the angle of focus of the camera 20º to 45º, the system correctly interprets 100% of the head moving towards or away. The novel detail is that the system allows entry and exit of the command mode by means of a special movement of the head, if the user enters the commands mode, the following options are available: tilt his head to the left, right, forward or backward.
Renjith V. Ravi, S.B. Goyal and Chawki Djeddi. Image Encryption using Quadrant Level Permutation and Chaotic Double Diffusion
Abstract: Information security is becoming more and more important as information technology advances. In the fields of telecommunication networks, diagnostic imaging, and multimedia applications, information security is crucial. The use of images as a medium for the conveyance of information is integral to the information age. However, image data differs from traditional text-based information because it is dense and has a strong pixel correlation. As a consequence, image encryption no longer works with traditional encryption techniques. Consequently, a safe and difficult-to-crack image encryption technique is required. This research suggests a new symmetric image encryption technique based on an improved Henon chaotic system with byte sequences and a unique method of shuffling a picture's pixels. This algorithm produces effective and efficient encryption of images. The suggested image encryption technique created a new dimension for safe image transfer in the realm of digital transmission by causing more confusion and diffusion, according to statistical analysis and experimental key sensitivity analysis.
Ahmed Mohamed Saad Emam Saad. Leveraging Graph Neural Networks for Botnet Detection
Abstract: Guarding the cyberinfrastructure is critical to ensure the proper transmission and availability of computer network services, information, and data. The proliferation in the number of cyber attacks launched on the cyberinfrastructure by making data unprocurable and network services inaccessible is on the rise. Botnets are considered one of the most sophisticated cybersecurity threats to the cyberinfrastructure and are becoming more daunting with time. Developing an efficient and robust botnet detection technique is a priority to ensure the security and reachability of the cyberinfrastructure. In this research, we introduce a solution and explore the use of a novel neural network architecture leveraging a graph-based learning approach, namely Graph Neural Network (GNN) for botnet detection. GNN was used to benefit from the unique architecture of botnets and to omit the feature engineering step of the machine learning pipeline as it is a costly and cumbersome process. Additionally, we report the effectiveness of different GNN variations in terms of detecting botnets to get an insight into the performance of each model. The ISCX-Bot-2014 dataset was used to create a graph data object for the training and testing of our proposed approach. The results show our proposed GNN solution’s ability to generalize to unseen botnets and perform better compared to other relevant work from the literature with an accuracy that exceeds 94%.
Shafina Abd Karim Ishigaki, Ajune Wanis Ismail, Nur Ameerah Abdul Halim and Norhaida Mohd Suaib. Voice Commands with Virtual Assistant in Mixed Reality Telepresence
Abstract: Mixed Reality (MR) telepresence is an expanding field in marketing research and practice especially in computer vision due to its ability to connect people in a remote location. As voice command and speech recognition technology have also been significant milestones in the twenty-first century, this technology has the advantage to be used in MR telepresence for interaction. Meanwhile, the virtual assistant avatar's implementation in MR telepresence can provide support for the interaction by reducing the workload and improving the performance of the interaction in collaborative MR telepresence. Therefore, in this paper, we propose to explore the voice commands interaction and describe the implementation with the virtual assistant avatar in MR telepresence. We implement speech-to-text (STT) conversion and utilize the semantics of Natural Language Processing (NLP) for the voice command. Meanwhile, for the avatar to respond, the speech service is used to convert the text-to-speech (TTS) input. This paper ends with a conclusion, limitations and suggestions for future works.
Abdullah Ahmed Aldulaimi, Alaa Ali Hameed and Akhtar Jamil. Photovoltaics Cell Anomaly Classification and Detection Using Deep Learning Techniques
Abstract: In recent years, solar photovoltaic (PV) systems have seen widespread use in the field of environmentally friendly energy harvesting. Additionally, the rate at which units reach the end of their useful life cycle is increasing. Heavy metals like lead, tin, and cadmium can be found in solar modules and cause environmental damage. Regular inspections and maintenance for your solar modules are essential to extending their useful life, cutting down on en-ergy waste, and keeping the environment safe. This thesis proposes a system that uses PV cell electroluminescence (EL) and deep learning techniques to efficiently screen for and categorize solar modules that exhibit anomalous behavior. Deep neural networks can accurately predict anomalies and classi-fy types of anomalies. Using PV cell electroluminescence, we propose con-volution neural network techniques based on residual network architecture and ensemble technology to accurately predict and classify anomalous solar modules.
Piyush Pant, S.B. Goyal, Anand Singh Rajawat, Amol Potgantwar, Pradeep Bedi and Chawki Chawki Djeddi. Anomaly Detection Algorithm with blockchain to prevent Potential Security Attacks in the IIoT Model of Industry 5.0
Abstract: No model is complete without security. Keeping that in mind, this research proposes to implement the Anomaly detection algorithm of Unsupervised Machine Learning on a model that is integrated with Blockchain. Machine Learning and Blockchain are growing technologies of this era with huge untapped potential. Exploration of these two domains would take existing technologies to new level like Internet of Things, Nanotechnology, Digital Twins, Healthcare, etc. The Blockchain technology enhanced the overall security of the Industrial Internet of Things (IIoT) model and upgraded it to industry 5.0. IIoT is one of the key technologies for upcoming industry as the interaction between humans and machine would increase significantly. However, to secure it even further and make the model intelligent, the machine learning algorithm is proposed. Anomaly detection algorithm with Gaussian distribution is proposed as it would monitor the working of the IIoT and detect if there is any unusual activity going on. The threshold for an activity to be classified as unusual or anomalous is discussed in the paper along with the difference between classification algorithm and anomaly detection algiorithm. The research focuses on the theory of the algorithm and its implementation using the latest tools and technology, it later discusses its integration with the model and Blockchain. The algorithm also focused on other applications of the model like detection of faulty driver, device, or equipment.
Md. Mijanur Rahman, Zohan Noor Hasan, Mahanaj Zaman Marufa and Mukta Roy. Covid-19 seasonal effect on infection cases and forecasting using Deep Learning.
Abstract: The COVID-19 pandemic infected billions of people worldwide. The government has taken a number of steps to control infection cases, but due to frequent changes in variants, it is difficult to control the infection rate, and taking precautions over a long period of time is infelicitous. Any correlation between the virus's ascendancy and changes in climate or temperature is difficult to detect. In this research, different countries, and seasons have been investigated to assess the relationship between temperature and the rate of infection. In predicting the infection rate of new cases, popular deep learning (DL) methods, long short-term memory (LSTM), and gated recurrent units (GRUs) have been applied here. Infection cases were visualized, including the time period from previous data. Individual countries have particular weather conditions that vary between countries; for this reason, the temperature has taken in a specified range in seasonal-based forecasting. To specify a particular temperature or season is challenging but combining season and temperature from past data generated a pattern. It shows that the highest number of infection cases reached at a certain time of the season and found a seasonal effect on COVID-19.
Abdulhameed Aboumadi and Hilal Arslan. PV Output Power Prediction Using Regression Methods
Abstract: In recent years, there has been a growing public awareness of climate change and its causes greenhouse gas emissions particularly CO2. Thus, individuals, businesses, and governments around the world have taken steps to reduce their emissions. One of these steps is to increase adoption of renewable energy sources, such as solar power, which has low building and maintenance costs and requires minimal mainte-nance. Accurate estimation of solar energy production is crucial to ensure the stabil-ity of electrical networks as the transition to renewable energy sources such as solar power increases. In this study, machine learning regression algorithms including artificial neural networks, support vector regression, regression trees, and k-nearest neighbor are performed to estimate hourly solar energy production of one month using historical production data including various meteorological parameters. The models are optimized using grid search. The performance of the models is evaluated using the RMSE, MAE, and R2 evaluation metrics. The results showed that the k-nearest neighbor regression model achieves the highest performance with an R2 score of 0.9715.
Maliha Tabussum, Murat Kuzlu, Ferhat Ozgur Catak, Salih Sarp and Kevser Şahinbaş. Secure Future Healthcare Applications Through Federated Learning Approaches
Abstract: The healthcare field is so sensitive regarding data privacy and security due to including medical and personal information. Almost all healthcare applications are required to increase data security and privacy, which use traditional machine learning approaches relying on centralized systems, both computing resources and the entirety of the data. Federated learning, a sort of machine learning technique, has been used to exactly address this issue. The training data is disseminated across numerous devices in federated learning, and the learning process is collaborative. There are numerous privacy attacks on Deep Learning (DL) models that attackers can use to obtain sensitive information. As a result, the DL model should be safeguarded from adversarial attacks, particularly in healthcare applications that use sensitive medical data. This paper provides a comprehensive review of federated learning on future healthcare applications. It also discusses the types of federated learning along with its implementation in healthcare applications.
Sevval Colak, Arezoo Sadeghzadeh and Md Baharul Islam. T-SignSys: An Efficient CNN-based Turkish Sign Language Recognition System
Abstract: Sign language (SL) is a communication tool playing a crucial role in facilitating the daily life of deaf or hearing-impaired people. Large varieties in the existing SLs and lack of interpretation knowledge in the general public lead to a communication barrier between the deaf and hearing communities. This issue has been addressed by automated sign language recognition (SLR) systems, mostly proposed for American Sign Language (ASL) with limited number of research studies on the other SLs. Consequently, this paper focuses on static Turkish Sign Language (TSL) recognition for its alphabets and digits by proposing an efficient novel Convolutional Neural Network (CNN) model. Our proposed CNN model comprises 9 layers, of which 6 layers are employed for feature extraction, and the remaining 3 layers are adopted for classification. The model is prevented from overfitting while dealing with small-scale datasets by benefiting from two regularization techniques: 1) ignoring a specified portion of neurons during training by applying a dropout layer, and 2) applying penalties during loss function optimization by employing L2 kernel regularizer in the convolution layers. The arrangement of the layers, learning rate, optimization technique, model hyper-parameters, and dropout layers are carefully adjusted so that the proposed CNN model can recognize both TSL alphabets and digits fast and accurately. The feasibility of our proposed T-SignSys is investigated through a comprehensive ablation study. Our model is evaluated on two datasets of TSL alphabets and digits with an accuracy of 97.85% and 99.52%, respectively, demonstrating its competitive performance despite straightforward implementation.
Abstract: This study uses machine learning methods to find deterioration in turbomachine parts. In turbomachines, damage control procedures are carried out at specific times. Even though these checks take a while, if there is no damage, the components won’t be replaced, and it is not anticipated that they will be rechecked until the following control or an unforeseen incident. For this situation, a machine learning algorithm has been developed and 96\% accuracy was obtained for overall components.
Piyush Pant, Anand Singh Rajawat, S. B. Goyal, Pawan Bhaladhare, Pradeep Bedi and Chawki Djeddi. Data analysis and clustering using K-means algorithm for Mars Data with the help of Perseverance Rover
Abstract: Mars is considered as one of the most potential habitable planets for humans. To discover more and answer curious questions, the exploration of the planet Mars must be continued and expanded using advanced tools and technologies. The Perseverance rover was sent to the Jezero crater on Mars to find the answers to such questions. This research describes the rovers of Mars with their objectives and explains the structure of the Perseverance rover in detail along with its objective and other related information. The research later discusses why the Jezero crater was chosen as the landing site and why the research proposes to take the help of the rover. The research covered the topics like rovers and landing sites because it was a crucial part of the research to understand the process, working, and implementation of the model. This research proposes to implement algorithms from the Unsupervised Machine Learning domain to develop a model that helps to use the data collected from mars to find some patterns to learn more about the surface of mars, spread awareness, and increase the interest of people in space research and exploration. The basic mathematical implementation of the K-means is described in the research that could be used for data analysis to find pattern or clusters in the data collected from rover. The paper also uses the Elbow method to find the number of clusters ‘k’. The power, stability, and flexibility of unsupervised machine learning to analyze complex data and find clusters using the K-Means clustering algorithm is shown in the research.
Sura Ali Hashim and Hasan Huseyin Balik. Deep Learning for ECG Signal Classification in Remote Healthcare Applications
Abstract: Due to several current medical applications, the significance of Electrocardiogram (ECG) classification has increased significantly. To evaluate and classify ECG data, a variety of machine learning methods are now available. Utiutilizing deep learning architectures, where the top layers operate as feature extractors and the bottom layers are completely coupled, is one of the solutions that has been suggested. In addition to classification results, this work also proposes a learning architecture for ECG classification utilizing 1D convolutional layers and Fully Convolution Network (FCN) layers. We made several changes to get the best result, getting 98% accuracy and 0.2% loss. A comparison has been made and showed that our work is better than other related works
Zuhair Alhous, Muhannad Al-Waily, Muhsin Jweeg and Ahmed Mouthanna. Computerize simulation of nonlinear vibration sandwich plate structure with porosity Functionally Graded Materials core
Abstract: This research presents a novel approximation accurate value of analysis of the nonlinear vibration to determine the frequency of sandwich plates with functionally graded sections and porosities. The kinematic relations are created and controlling differential equation by making use of the first ordinal differential shear deformation theory. It is presumed that the FGM plate is formed of an isotropic material that has a porosity distribution throughout its surface. Under the power-law scheme, the only direction in which the qualities of the material fluctuate smoothly is in the direction of thickness. Analyses are done to determine how the nonlinear vibration parameters of functionally graded sandwich plates are affected when variables such as the gradient indices, porosity distribution, boundary conditions, and geometrical attributes are altered. Employing the FOSD theory, researchers conduct a thorough numerical examination. The outcomes with various boundary conditions demonstrate how the porosity distribution affects the nonlinear vibration properties of FG sandwich plates. The outcome showed that the approximation approach and FOSD theory have a fair level of agreement.
Sumeia Mechi, Muhsin Jweeg and Muhannad Al-Waily. The Computer Modelling of the Human Gait Cycle for the Determination of Pressure Distribution and Ground Reaction Force Using a Below Knee Sockets
Abstract: In this research, many composite configurations were presented employing person, kevlar, and carbon fibres in addition to the kenaf fibres, which were used to create the original below-knee (B.K.) prosthesis. These modifications were made for convenience and to lengthen the useful life of our prosthesis. The study used a layered experimental design, with some layers containing kenaf and others without. The ideal strength-to-weight ratio (E/) and the desired modulus of elasticity were sought, together with the effective lamination employed in the production of sockets. They prepared the socket produced from natural kenaf fibres and analyzed its performance throughout the gait cycle comprising the experimental phase. A kenaf prosthetic socket was provided to a 58-year-old man who lost his right Leg to diabetes and weighed 80 kg. This socket took group D plugs and jacks. Tests of the ground reaction force (G.R.F.) showed that the difference between the stance and swing phases did not exceed (-5.3%), or 5.3% improvement over a different instance, signifying a 20.8% improvement.
Sanabil Mahmood and Monji Herallah. Evaluation of the Existing Web Real Time Signalizing Mechanism for Peer-to-Peer Interconnecting: Survey
Abstract: Web Real Time Interconnecting (Web.RTC) is prepared to allowance the co-occurrence of audio, video, and data interconnectings. Also, it is a set of standards, libraries, and JavaScript APIs. Web.RTC had a number of advantages, including the lack of plug-ins, the ease of usage, no licensing, and the excellent quality of the RTC applications. However, signalizing mechanism that setup, establish and end a interconnecting between peers has not specified in Web.RTC. This paper reviews general studies and methods that used and suggested for Web.RTC signalizing protocols. Moreover, it focuses on the limitations on multi- web crawlers interconnecting, the network topology for exchanging data and multimedia and leveraging public groundwork to manage privacy as a service or information, or public architecture to handle signalizing protocols. Therefore, this work focuses on researching in related work for the existing Web.RTC signalizing mechanisms/protocols in order to find out the limitations and the main gab at this time and also for a thorough knowledge of signalizing in Web.RTC including their advantages and disadvantages.
Fazliaty Edora Fadzli, Ajune Wanis Ismail, Norhaida Mohd Suaib and Yin Yee Lau. Autonomous Agent Using AI Q-Learning in Handheld Augmented Reality Ludo Board Game
Abstract: An autonomous agent works with Artificial Intelligence (AI) can decide its actions to adapt and respond to the changes in a dynamic environment. The autonomous agent can be developed in games as a Non-Player Character (NPC) to interact with the changes of state in the game environment. Tradi-tional board games such as Ludo have had many players since the olden days but slowly lost attraction to the public, especially the younger genera-tions as digital games become more popular. Although the Ludo board game can be digitized to fascinate the players through implementing Augmented Reality (AR) technology in handheld devices, common NPCs found in games have determined actions and are unable to learn from experience and adapt to the changes of the game environment. Therefore, this research aims to develop an autonomous agent for board game in handheld AR (HAR). The first step in the three main phases is to examine the autonomous agent for the HAR board game. The second phase is developing the AR board game with Q-learning and Minimax algorithms for board game agents. Final-ly, the third phase is integrating the AR board game with Q-learning and Min-imax agents in handheld. The novel contribution of this research is the rede-sign of Ludo for AR with autonomous agent and generate the training data using the Q-Learning algorithm to create autonomous agent in AR.
Senthamarai Kannan K and Parimyndhan V. Modelling and Estimating of VaR through the GARCH model
Abstract: This study focuses on the analysis of fiscal series with time-varying conditional variance utilizing the ARIMA-GARCH with Value at Risk (VaR) model. ARIMA-GARCH can predict risk when stock variance is Heteroscedasticity. The price of the Reliance stock is analyzed for fifty months. This research findings indicate that the VaR is a useful technique to reduce risk exposure and perhaps avoid losses when investing in the Reliance stock. The findings show that ARIMA (0,0,0)-GARCH (1,1) has the best fit, with an Akaike information criterion (AIC) value of -5915.325, at a confidence level of 95%. The GARCH technique is used to determine the conditional variance of the residuals and contrasts it with the delta-normal method. At a 95% confidence level, the VaR is used to calculate the likelihood of losing an investment by 2.7% or more in a single day.
Serdar Yildiz, Abbas Memiş and Songül Varlı. Nuclei Instance Segmentation in Colon Histology Images with YOLOv7
Abstract: In histology image analysis, instance-based nuclei segmentation is one of the challenging tasks within the segmentation-guided studies since it is quite troublesome to detect each distinct nuclei instance of each nuclei type in images in contrast to the semantic segmentation in which all the image pixels of a nuclei type are labeled with the same mask ID although the segmented region may comprise of multiple instances. In this paper, an instance-based medical image segmentation task is addressed, and in this context, instances of multiple types of nuclei in colon histology images are aimed to be delineated distinctly. For the instance-based segmentation of the nuclei in colon histology images, the YOLOv7 algorithm and its built-in instance segmentation module are utilized. In the experimental studies performed on Colon Nuclei Identification and Counting (CoNIC) Challenge 2022 colon histology image dataset by using a 5-fold cross validation performance evaluation strategy, nuclei instances belonging to 6 classes as the neutrophil, epithelial, lymphocyte, plasma, eosinophil and connective were segmented. To calculate the overall system accuracy, the quantification metrics of mean average precision (mAP) and mean panoptic quality (mPQ) were measured. In performance evaluations, quite promising accuracy values were obtained. The mAP values of 0.2885 and 0.2903, and mPQ values of 0.1659 and 0.1704 were observed by using the YOLOv7 algorithm. To the best of our knowledge, this is the first nuclei instance segmentation study with YOLOv7.
Yusuf Şevki Günaydın and Ömer Mintemur. Evolutionary Approach to Feature Elimination in House Price Estimation
Abstract: One of the most basic human needs is the need for shelter. Since ancient times, people have been looking for a house that is both safe and affordable. However, in modern times, although safety is no longer an issue, the definition of an affordable house has changed. In the past, affordability did not depend on many parameters as it does today. However, today, this definition depends on different features, such as the location of the house, the year of construction, the number of rooms, etc. These features affect the level of affordability, and consequently the price of the house. Since houses are also used as an investment option, correct estimation of house prices is an important issue. The determination of features that have a significant impact on the price of a house is a subjective notion and, therefore, requires an objective approach. Thanks to technological developments, Artificial Intelligence algorithms remove the human factor in most of the decision-making processes. In this study, a naive approach was proposed to estimate house prices by selecting the most effective features of a house. To select the most effective features, Genetic Algorithm approach was utilized. For estimation, LightGBM was used. The AmesHouse data set was used for the experiments. The results suggested that the proposed method both reduced the features and produced lower estimation errors than other proposed methods that used the same dataset.
Emre Dandıl. A Hybrid Machine Learning Approach for Brain Tumor Classification using Artificial Neural Network and Particle Swarm Optimization
Abstract: Brain tumors have an increasing trend in recent years and are one of the main causes of death. Therefore, computer-assisted secondary tools that can help diagnose brain tumors at an early stage are needed. It is crucial to use machine learning methods which can help classification of brain tumors. In this paper, a hybrid machine learning approach is proposed for classification of brain tumors using artificial neural network and particle swarm optimization on magnetic resonance (MR) images. The approach is composed of six steps. The first step includes enhancement of MR images and the second step consists of eliminating the skull region. The third step is composed of extracting the region of interest through segmentation of the masses with s-FCM method. In the fourth step, feature extraction of the segmented tumors is undertaken with four different methods and feature selection is applied with relief-f and sequential floating forward selection (SFFS) methods in fifth step. In the last step, benign and malignant tumors are classified with Bayes, support vector machines (SVM), artificial neural network (ANN) and particle swarm optimization-based artificial neural network (PSO-ANN) classification methods and compared with each other. According to the experimental studies, the proposed approach provides high scores such as 96.28% accuracy, 97.58% sensitivity, 93.75% specifity values using PSO-ANN. As a result, the proposed approach can facilitate the decision making of radiologists for brain tumor classification.
Hazrat Bilal and Cem Direkoğlu. Optimized KiU-Net: Lightweight Convolutional Neural Network for Retinal Vessel Segmentation in Medical Images
Abstract: Medical image segmentation helps with computer-assisted disease analysis, operations, and therapy. Blood vessel segmentation is very important for the diagnosis and treatment of different diseases. Lately, the U-Net and KiU-Net based vessel segmentation techniques have demonstrated reasonable achievements. The U-Net architecture belongs to the group of undercomplete autoencoders which ignores the semantic features of the thin and low contrast vessels. On the other hand, the KiU-Net uses a combination of undercomplete and overcomplete architectures to segment the small structure and fine edges better than U-Net. However, this solution is still not accurate enough and computationally complex. We propose an Optimized KiU-Net model to increase the segmentation accuracy of thin and low-contrast blood vessels and improve the computational efficiency of this lightweight network. The proposed model selects the ideal length of the encoder and the number of convolutional channels. Moreover, our proposed model has better convergence and uses a smaller number of parameters by combining the feature map at the final layer instead at each block. Our proposed network outperforms the KiU-Net on vessel segmentation in the RITE dataset. It obtained an overall enhancement of about 4% in terms of F1 score and 6% in terms of IoU compared to KiU-Net. Evaluation and comparison were also conducted on the GLASS dataset, and the results show that the proposed model is effective.
Samer Nofal, Amani Abu Jabal, Abdullah Alfarrarjeh and Ismail Hababeh. On the Characteristic Functions in Listing Stable Arguments
Abstract: An abstract argumentation framework ({\sc af}) is viewed as a directed graph such that graph vertices represent abstract arguments while graph edges denote attacks between these arguments. We say that a set, $S$, of arguments, are conflict-free if and only if for every $(x,y) \in S \times S$, $x$ does not attack $y$. A set, $S$, of arguments of a given {\sc af}, is called a stable extension in {\sc af} if $S$ is conflict-free and such that every argument outside $S$ is attacked by an argument inside $S$. To the best of our knowledge, a thorough mathematical analysis of the truth of what so-called characteristic functions (which are an essential component in generating all stable extensions of a given {\sc af}) is not previously addressed in the literature. We fill this gap; we rigorously analyze the verity of characteristic functions employed in listing all stable extensions in a given~{\sc af}.
Srinath K S, Kiran K, P Deepa Shenoy and Venugopal K R. Generating Sub-emotions from Social Media Data using NLP to Ascertain Mental illness
Abstract: It is predicted that mental illness will be one of the leading causes of death in 2030. Many people will not share their details of the illness detail with others, in-cluding family and friends. Also, many are unaware that their mental disorder is affecting their thinking and behavior. Early detection and medical intervention are necessary, otherwise it leads to severe problems. More than half of the world population, that is around 58.4% of the people use Social Media (SM) to express their thoughts and feelings. By fetching their timely thoughts and feelings ex-pressed in social media we can analyze their emotions and sub-emotions. In this study, we developed a novel model to generate the sub-emotions of social media users from EmoLEX lexicon using the Affinity Propagation (AP) algorithm and word2vect conversion-word2vec-google-news-300. The number of clusters and vocabulary obtained is evaluated by using word2vect conversions. It is found that, by using AP algorithm the consistency of words are equally distributed in each cluster with respect to all emotions.
Kevser Şahinbaş,Ferhat Ozgur Catak, Murat Kuzlu, Maliha Tabussum and Salih Sarp. Non-Cryptographic Privacy Preserving Machine Learning Methods: A Review
Abstract: In recent years, the purpose of Machine Learning (ML) techniques exploiting data to produce predictive models is widely emerging in decision-making and solving specific problems for various fields, including healthcare, energy, retail, transportation, and many more. Generally, a well-performing ML model is based on large volumes of training data. On the other hand, collecting data and using it to predict behavior poses significant challenges to the privacy of individuals and organizations, such as data breaches, loss of privacy, and corresponding financial damage. Therefore, well-designed privacy-preserving ML (PPML) methods are significantly required for many emerging applications to mitigate these problems. This paper presents a comprehensive study of non-cryptographic privacy-preserving ML along with the selected methods, i.e., differential privacy and federated learning. This paper is expected to provide a roadmap for future research directions in the PPML field.
Mübarek Mazhar Çakır and Gökalp Çınarer. Detection And Comparative Results Of Plant Diseases Based On Deep Learning
Abstract: Plant diseases are one of the problems that threaten crop health and yield in agriculture. Various diseases occurring in plants harm human health and economically producers and producer countries. Early diagnosis is very important in order to prevent the damage caused by diseases. For the early detection of these diseases in plants, continuous observation and examination of plants is required. In large agricultural areas, continuous monitoring of the plants by the producers or workers requires long periods of time and causes extra cost increase. In addition, the person who studies plant leaves must be an expert in plant science. A study was carried out to detect diseases by observing plants based on deep learning, which will be a technological solution to all these problems. Yolov5 and Yolov6 algorithms, one of the object recognition algorithms, was used for plant disease diagnosis. After comparing the two algo-rithms, the highest AP value with 58.4% belongs to the Yolov5-m model, the highest AR value with 69.3% belongs to the Yolov6-s model, and the highest F1 score with 62.4% belongs to the Yolov5-m. With the study, the comparative results of the models of the Yolo algorithms, together with the hyperparameter values, are given. According to the obtained values, it is seen that the small size models give the best performance. The higher performance of the small size models shows that deep learning models can be integrated into a mobile system, enabling rapid plant identification, sustainability in agriculture and cost reduction.
Ayla Gülcü, İsmail Taha Samed Özkan, Zeki Kuş and Osman Furkan Karakuş. Triplet MAML for few-shot classification problems
Abstract: In this study, we propose a TripletMAML algorithm as an extension to Model-Agnostic Meta-Learning (MAML) which is the most widely-used optimization-based meta-learning algorithm. We approach MAML from a metric-learning perspective and train it using meta-learning tasks composed of triplets of images. The idea of meta-learning is preserved while generating the meta-learning tasks and training our novel meta-model. The experimental results obtained on four few-shot classification datasets show that TripletMAML that is trained using a combined loss yields in high quality results. We compared the performance of TripletMAML to several metric learning-based methods and a baseline method, in addition to MAML. For fair comparison, we used the reported results of those algorithms that were obtained using the same shallow backbone. The results show that TripletMAML improves MAML by a large margin, and yields better results than most of the compared algorithms in both 1-shot and 5-shot settings. Moreover, when we consider the classification performance of other meta-learning algorithms that use much deeper backbones, we conclude that TripletMAML is not only competitive in terms of the classification performance but also very efficient in terms of the complexity.
Himmet Toprak Kesgin and Mehmet Fatih Amasyali. Iterative Mask Filling: An Effective Text Augmentation Method Using Masked Language Modeling
Abstract: Data augmentation is an effective technique for improving the performance of machine learning models. However, it has not been explored as extensively in natural language processing (NLP) as it has in computer vision. In this paper, we propose a novel text augmentation method that leverages the Fill-Mask feature of the transformer-based BERT model. Our method involves iteratively masking words in a sentence and replacing them with language model predictions. We have tested our proposed method on various NLP tasks and found it to be effective in many cases. Our results are presented along with a comparison to existing augmentation methods.
Şerif İnanır and Yılmaz Kemal Yüce. Sub Data Path Filtering Protocol for Subscription of Event Parts and Event Regeneration by Broker System in Pub/Sub Pattern
Abstract: Pub/Sub is a common pattern allowing a producer to publish events to consumers. In types of Pub/Sub, structure of an event is either identified by publishers based on static rules or by consumers based on filtering approaches. In both scenarios, actors’ total performance might get degraded due to required operations (e.g., filtering) impacting throughput. This study focuses on designing a filtering approach for both actors of Pub/Sub by reducing data size to be transmitted by producers and received and processed by consumers by creating a loosely coupled context, in which horizontal alterations to structure of any event can occur. Sub Data Path (SDP) approach presents a matching tree to separate an event with scope like JSON data, and each key in the relevant event act like a topic without being defined as a topic. Thereby, producers only must transmit part of a message through a path on the event structure to be located into a former event to create new event; consumers can subscribe to any subtopic (key for JSON format) to be able to receive data in terms of its own mechanism, not a producer’s design. Therefore, creating an event can be completed with different producers which contribute a piece of the whole event; Bounded Context structure belongs to microservice architecture as a decomposition strategy can be handled by consumers in relation to their own business logic. To measure the proposed method, an experiment with gaze points collected by an eye tracker has been designed. By performing the filtering method for one, two and maximum SDP keys, filtering duration, event size reduction percent and transmission duration were revealed. The experimental results imply that the proposed method can send 7.5 events in average, instead of sending just one in the same period. Also, since worst case of the proposed method basedon events in the context can be calculated, an architecture can be prevented from bottlenecks. These benefits makes SDP advantageous over similar methods in terms of being both a fast and scalable alternative.
Muhannad Al-Waily and Muhsin Jweeg. A computerized Experimental Rig for the Vibration Investigation of Cracked Composite Materials Plate Structures
Abstract: A composite plate is strengthened by reinforcements in the form of powder particles, and short or long fibers, where, the fiber glass was used as the reinforcement fiber to modifying the composite plate was investigated. The crack effects on the plate natural frequencies were investigated using different boundary conditions (SSSS, SSCC, SSFF). The problem was solved numerically using the Finite Element Method, adopting the ANSYS program, experimentally employing the time variation and measuring the natural frequency. The comparison between the finite element method and the experimental program has shown good agreement with a maximum discrepancy of not more than 8.5% for the tested cases of the composite plates. It was noticed that the natural frequency decreases with the presence of the crack. It was also noticed that the central damage effect has a higher effect on the natural frequency than the other crack positions. Also, the results shows that the natural frequency decreases as the crack length or depth increases due to the stiffness reduction in cracked samples. And, the increase in the aspect ratio of the plate means an increase in the mass of the plate, which results in a decrease in the natural frequencies. In addition, the results indicate a decrease in the natural frequency for the SSCC boundary condition was less than the decrease in the natural frequency for other boundary conditions, SSSS and SSFF. Finally, the results showed that the short fiber types modifying of the natural frequency more than the modifying for natural frequency of other reinforcement fiber types