Browsing by Author "Talagala, PD"

Now showing 1 - 20 of 30

item: Article-Full-text
Anomaly detection in high-dimensional data
(Taylor and Francis, 2021) Talagala, PD; Hyndman, RJ; Miles, KM
The HDoutliers algorithm is a powerful unsupervised algorithm for detecting anomalies in high-dimensional data, with a strong theoretical foundation. However, it suffers from some limitations that significantly hinder its performance level, under certain circumstances. In this article, we propose an algorithm that addresses these limitations. We define an anomaly as an observation where its k-nearest neighbor distance with the maximum gap is significantly different from what we would expect if the distribution of k-nearest neighbors with the maximum gap is in the maximum domain of attraction of the Gumbel distribution. An approach based on extreme value theory is used for the anomalous threshold calculation. Using various synthetic and real datasets, we demonstrate the wide applicability and usefulness of our algorithm, which we call the stray algorithm. We also demonstrate how this algorithm can assist in detecting anomalies present in other data structures using feature engineering. We show the situations where the stray algorithm outperforms the HDoutliers algorithm both in accuracy and computational time. This framework is implemented in the open source R package stray. Supplementary materials for this article are available online
item: Conference-Full-text
Arima and ann approach for forecasting daily stock price fluctuations of industries in Colombo stock exchange, Sri Lanka
(Faculty of Information Technology, University of Moratuwa., 2020-12) Wijesinghe, GWRI; Rathnayaka, RMKT; Karunananda, AS; Talagala, PD
Time series forecasting is regarded as the most successful criterion among several factors involved in the decision-making process to pick a correct prediction model. Improving predictability has become crucial for decision-makers and managers, especially time series forecasts, in various fields of science. Using K-mean clustering and Principle Component Analysis, the dataset is clustered based upon a central point selection and the Euclidian distance measurement. The results define the main contribution sector for CSE, and the business in the selected sector in the 2008-2017 period in accordance with the clustering results. In particular, ARIMA has demonstrated its performance in predicting the next lags in precision and accuracy. With regard to Colombo Stock Exchange (CSE), there are very few studies in the literature that have focused on new approaches to forecasts of high volatility stock price indexes. Different statistical methods and economic data techniques have been widely applied in the last decade in order to classify CSE's stock price, patterns and trade volumes. This article looks at the best sector and organization to invest in and discusses whether and how the deep-learning algorithms for time series data projection, such as the Back Propagation Neural Network, are better than traditional algorithms. The results show that Deep learning algorithms like BPNN outperform traditionally based algorithms like the model ARIMA. For ARIMA and ANN, MAPE values are 0.472206 and 0.1783333 respectively. MAE values are 29.6975 and 4.708423 respectively results for ARIMA and ANN. The MAE and MAPE values relative to ARIMA and BPNN, which suggests BPNN `s superiority to ARIMA.
item: Conference-Full-text
Artificial neural network to estimate the paddy yield prediction using remote sensing, weather and non weather variable in Ampara district, Sri Lanka
(Faculty of Information Technology, University of Moratuwa., 2020-12) Wanninayaka, WMRK; Rathnayaka, RMKT; Udayakumara, EPN; Karunananda, AS; Talagala, PD
In Sri Lanka, seasonal paddy area mapping and rice prediction is based on the traditional methods with poor technologies. Ampara district has been chosen as the study area because its contribution is considered as the second highest paddy yield to the Sri Lankan rice harvest. This study focuses on developing models for precise mapping paddy and predicting the harvest of rice in the Ampara district. It helps the government and persons of authority to take decisions about how to manage the economy based on the rice quantity. Research includes the imageries of satellites sentinel-1 and sentinel-2 the period from April to September 2019. The two classification methods, Divisional Secretory Division (DSD) and maximum likelihood classification were used to identify the real paddy area. The accuracy rates of these classifications were 0.92 and 0.86 respectively. Artificial Neural Network (ANN) model was used to predict paddy rice harvest using sentinel 2 features extracts and round truth data. Mean square error of the model is 0.106 and mean absolute error is 0.245. Increasing the remote sensing imagery directly affects to enhance accuracy. Increasing the number of sample classes and number of classes in various types will raise-up higher accuracy than in here.
item: Conference-Full-text
An automated decision-making framework for precipitation-related workflows
(Faculty of Information Technology, University of Moratuwa., 2020-12) Adikari, AMH; Bandara, HMND; Herath, S; Chitraranjan, C; Karunananda, AS; Talagala, PD
Due to weather’s chaotic nature, static workflow managers are ineffective in integrating multiple Numerical Weather Models (NWMs) with cascading relationships. Unexpected events like flash floods and breakdown in canal water control systems or reservoirs make decision-making in workflow management further complicated. To enable dynamic decision-making, we need to update part or entire workflow, terminate unfitting NWM executions, and trigger parallel NWM workflows based on recent results from NWMs and observed conditions. Most of the existing weather-related decision support systems cannot trigger or create workflows dynamically. They are also designed for specific geography or functionality, making it challenging to customize for regions with different weather patterns. In this paper, we present an automated decision-making framework for precipitation-related workflows. The proposed framework can manage complex weather-related workflows dynamically in response to varying weather conditions, automatically control and monitor those workflows, and update workflow paths in response to unexpected weather events. Using significant flood-related datasets from the Colombo catchment area, we demonstrate that the proposed framework can achieve 100% accuracy in dynamic workflow generation and path updates compared to manual workflow controlling. Also, we demonstrate that unexpected event identification and pumping station controlling workflow triggers could be improved with advance rule sets.
item: Conference-Full-text
Automatic diagnosis of diabetic retinopathy using machine learning: a review
(Faculty of Information Technology, University of Moratuwa., 2020-12) Gunawardhana, PL; Jayathilake, R; Withanage, Y; Ganegoda, GU; Karunananda, AS; Talagala, PD
Diabetic Retinopathy is a popular cause of diabetes, causing vision-impacting lesions of the retina. Blindness may be avoided by early detection. The ophthalmologist's manual approach of diagnosing diabetic retinopathy is expensive and time consuming. At the same time, unlike computer assisted diagnostic systems, it may cause misdiagnosis. Deep learning has recently become one of the most effective approaches that has obtained better efficiency in the analysis and classification of medical images. In medical image analysis, convolutional neural networks are more commonly used as a deep learning approach and they are extremely effective. This paper assessed and addressed the new state-of-the-art Diabetic Retinopathy color fundus image classification and detection methodologies using deep learning and machine learning techniques. Additionally, various challenging issues that need further study are also discussed.
item: Conference-Full-text
Btdm: a qos-based trust distribution mechanism for cloud computing
(Faculty of Information Technology, University of Moratuwa., 2020-12) Firdhous, MFM; Budiarto, R; Karunananda, AS; Talagala, PD
Cloud computing makes the delivery of computing resources over the Internet as services. As there are many providers in the market, it is necessary to monitor their performance. Several mechanisms for monitoring service quality of providers have been reported in the literature. But, it is not possible to monitor the entire cloud system by a single monitor. Hence, there is a need for a mechanism to share the performance metrics across a large geographical area. In this paper, the authors propose a mechanism called Bayesian Trust Distribution Mechanism (BTDM) for sharing the performance metrics as trust scores across an extended geographical area. The proposed BTDM also checks the reliability of the received scores based on their previous experience and adjusts them based on the reliability of sender. BTDM was tested using simulations and the results show that it performs better than the other mechanisms reported in the literature.
item: Conference-Full-text
Building social resilience during disasters: an investigation into the role of online social media networks
(Faculty of Information Technology, University of Moratuwa., 2020-12) Firdhous, MFM; Karunananda, AS; Talagala, PD
Within the last ten years, the world witnessed four serious epidemics. COVID-19 has been the most serious of these ones in terms of the number of people affected and the lives lost. In order to contain the spread of the disease many countries including Sri Lanka enforced 24 hour curfews. The social isolation created by lockdowns creates many problems in people including anxiety and depression. Many studies have been carried out on effect of lockdowns on mental well being of people. But, so far nobody has studied whether online social me can help people overcome the negative effects of lockdowns. This research was carried out to fill this gap. An online survey was carried out to understand how people used social media during the continuous curfew enforced by the Sri Lankan government. The research found that the average time spent using social media has increased compared to normal days. Also, majority of the users agreed that the social media helped them overcome the boredom created by the lockdown. This fact was confirmed using statistical tests in this study.
item: Conference-Full-text
Cross-vit: cross-attention vision transformer for image duplicate detection
(Information Technology Research Unit, Faculty of Information Technology, University of Moratuwa., 2023-12-07) Chandrasiri, MDN; Talagala, PD; Piyatilake, ITS; Thalagala, PD; Ganegoda, GU; Thanuja, ALARR; Dharmarathna, P
Duplicate detection in image databases has immense significance across diverse domains. Its utility transcends specific applications, adapting seamlessly to a range of use cases, either as a standalone process or an integrated component within broader workflows. This study explores cutting-edge vision transformer architecture to revolutionize feature extraction in the context of duplicate image identification. Our proposed framework combines the conventional transformer architecture with a groundbreaking cross-attention layer developed specifically for this study. This unique cross-attention transformer processes pairs of images as input, enabling intricate cross-attention operations that delve into the interconnections and relationships between the distinct features in the two images. Through meticulous iterations of Cross-ViT, we assess the ranking capabilities of each version, highlighting the vital role played by the integrated cross-attention layer between transformer blocks. Our research culminates in recommending a final optimal model that capitalizes on the synergies between higher-dimensional hidden embeddings and mid-size ViT variations, thereby optimizing image pair ranking. In conclusion, this study unveils the immense potential of the vision transformer and its novel cross-attention layer in the domain of duplicate image detection. The performance of the proposed framework was assessed through a comprehensive comparative evaluation against baseline CNN models using various benchmark datasets. This evaluation further underscores the transformative power of our approach. Notably, our innovation in this study lies not in the introduction of new feature extraction methods but in the introduction of a novel cross-attention layer between transformer blocks grounded in the scaled dot-product attention mechanism.
item: Conference-Full-text
Data mining approach for analyzing factors influencing vegetable prices
(Faculty of Information Technology, University of Moratuwa., 2020-12) Illankoon, IMGL; Kumara, BTGS; Karunananda, AS; Talagala, PD
Vegetables have a special place in the Sri Lankan economy. The price of vegetables, unlike the prices of other products, changes daily. There are several reasons for this and the examples include environmental conditions, supply variability, demand, festivals and seasonality, social environment, political conditions, etc. The main purpose of this research is to analyze and predict the factors influencing daily vegetable price fluctuations using data mining techniques. In this research, the most influential factors for vegetable prices were classified using the classification algorithms J48, Random Tree, Random Forest, and Support Vector Machine, taking into account the data obtained from the secondary data sources. The highest classification accuracy of 97.7143% was given by the Random Forest algorithm and it also recorded the best values for Precision, Recall, F-measure, and MCC comparing with the other three. Furthermore, it is clear that the Random forest algorithm is the most suitable to predict influential factors and it can be recommended for the purpose.
item: Conference-Full-text
Design and evaluation of a capacitive sensor for real time monitoring of gravimetric moisture content in soil
(Faculty of Information Technology, University of Moratuwa., 2020-12) Pieris, TPD; Chathuranga, KVDS; Karunananda, AS; Talagala, PD
Soil moisture content measurements are required widely in the fields of lawn maintenance, irrigation systems of farming, soil processing in civil engineering etc. The paper presents a novel two terminal sensor for real-time volumetric soil moisture content measurement. The sensor was designed and tested with two terminals and an epoxy layer. This sensor was designed such that the range of the sensor in the range of the moisture measurement. An a-stable multi vibrator-based circuit was used to measure the capacitance across the two terminals when inserted into soil. A microcontroller-based system was used to measure, calculate the volumetric soil moisture content and save to a SD card. The results validate the effective applicability of the developed sensor in volumetric soil moisture content measurement.
item: Conference-Full-text
Early identification of deforestation using anomaly detection
(Information Technology Research Unit, Faculty of Information Technology, University of Moratuwa., 2023-12-07) Wijesinghe, N; Perera, R; Sellahewa, N; Talagala, PD; Piyatilake, ITS; Thalagala, PD; Ganegoda, GU; Thanuja, ALARR; Dharmarathna, P
Research involving anomaly detection in image streams has seen growth through the years, given the proliferation of high-quality image data in various applications. One such application that is in urgent need of attention is deforestation. Detecting anomalies in this context, however, remains challenging due to the irregular and low-probability nature of deforestation events. This study introduces two anomaly detection frameworks utilizing machine learning and deep learning for the early detection of deforestation activities in image streams. Furthermore, Explainable AI was used to explain the black box models of the deep learning-based anomaly detection framework. The class imbalance problem, the inter-dependency between the images with time, the lack of available labelled images, a datadriven anomalous threshold, and the trade-off of accuracy while increasing interpretability in the black box optimization methods are some key aspects considered in the model-building process. Our novel framework for anomaly detection in image streams underwent rigorous evaluation using a range of datasets that included synthetic and real-world data, notably datasets related to Amazon’s forest coverage. The objective of this evaluation was to detect occurrences of deforestation in the Amazon. Several metrics were used to evaluate the performance of the proposed framework.
item: Conference-Full-text
An elephant detection system to prevent human-elephant conflict and tracking of elephant using deep learning
(Faculty of Information Technology, University of Moratuwa., 2020-12) Premarathna, KSP; Rathnayaka, RMKT; Charles, J; Karunananda, AS; Talagala, PD
Human settlement is spreading to forest boundary areas because of the population growth, it triggers disputes between elephants and humans, leading to the loss of property and life. Continuous monitoring and tracking of elephants are difficult due to their large size and movement. Therefore, large-scale for real-time detection and alert of elephant intrusion into human settlements, monitoring is needed. Many methods had been implemented for the elephant’s intrusion detection and warning systems. Wildlife conservation and the management of human-elephant conflict require a cost-effective method of monitoring elephant behavior. In this paper, a method for the identification of the elephant as an object using image processing is proposed. The major aim of the study is to minimize the human-elephant conflict in the forest border areas and the conservation of elephants from human activities as well as protect human lives from elephant attacks. We used a data set containing elephants and we developed an approach to distinguish elephants and other animals. We used the Convolutional Neural Network and achieved a maximum accuracy of 94 percent. The proposed method outperformed existing approaches and robustly and accurately detected elephants. It thus can form the basis for a future automated early warning system for elephants.
item: Conference-Full-text
Energy and power consumption analysis of a wireless sensor node without a voltage regulator
(Faculty of Information Technology, University of Moratuwa., 2020-12) Pieris, TPD; Chathuranga, KVDS; Kulasekera, AL; Guha, P; Mukhija, P; Karunananda, AS; Talagala, PD
Wireless sensor nodes are used in a wide range of areas such as environment monitoring, health monitoring, military and engineering applications to transfer sensor data from one location to another location. These sensor nodes usually have sensors, a microprocessor, transceiver, and a limited power supply. Most sensor nodes are configured to be in sleep mode and wakes up periodically and transfers the sensor data. In the sleep mode they consume less energy and most of the power is consumed in the wake-up time. The power consumption directly affects the life span of the sensor node. Sensor nodes typically use linear voltage regulators rather than switching ones to prevent switching noise, switching voltage ripples and to limit the footprint of the switching circuitry. Low dropout regulators are also used as they consume less energy than standard regulators and saves the battery energy. Even the low dropout regulators consume considerable amount of power at the high battery voltages and considerably high current flows through them. If there is way to save this energy, the lifespan of the sensor node can be extended. One way is avoiding the voltage regulator completely. Modern electronic components such as microcontrollers, sensors and transceivers can work in a wide range of voltages and they have internal voltage references. Therefore, the design of a sensor of a sensor node without voltage regulator is possible. In this paper we implemented a sensor node design without a voltage regulator, and we have evaluated and concluded that this design has up-to 40% energy saving compared with same sensor node design with a voltage regulator.
item: Conference-Full-text
Evaluation of re-identification risks in data anonymization techniques based on population uniqueness
(Faculty of Information Technology, University of Moratuwa., 2020-12) Bandara, PLMK; Bandara, HMND; Fernando, S; Karunananda, AS; Talagala, PD
With the increasing appetite for publicly available personal data for various analytics and decision making, due care must be taken to preserve the privacy of data subjects before any disclosure of data. Though many data anonymization techniques are available, there is no holistic understanding of their risk of re-identification and the conditions under which they could be applied. Therefore, it is imperative to study the risk of re-identification of anonymization techniques across different types of datasets. In this paper, we assess the re-identification risk of four popular anonymization techniques against four different datasets. We use population uniqueness to evaluate the risk of re-identification. As per the analysis, k-anonymity shows the lowest re-identification risk for unbiased samples of the population datasets. Moreover, our findings also emphasize that the risk assessment methodology should depend on the chosen dataset. Furthermore, for the datasets with higher linkability, the risk of re-identification measured using the uniqueness is much lower than the real risk of re-identification.
item: Conference-Full-text
Human activity recognition using cnn & lstm
(Faculty of Information Technology, University of Moratuwa., 2020-12) Shiranthika, C; Premakumara, N; Chiu, HL; Samani, H; Shyalika, C; Yang, CY; Karunananda, AS; Karunananda, AS; Talagala, PD
In identifying objects, understanding the world, analyzing time series and predicting future sequences, the recent developments in Artificial Intelligence (AI) have made human beings more inclined towards novel research goals. There is a growing interest in Recurrent Neural Networks (RNN) by AI researchers today, which includes major applications in the fields of speech recognition, language modeling, video processing and time series analysis. Recognition of Human Behavior or the Human Activity Recognition (HAR) is one of the difficult issues in this wonderful AI field that seeks answers. As an assistive technology combined with innovations such as the Internet of Things (IoT), it can be primarily used for eldercare and childcare. HAR also covers a broad variety of real-life applications, ranging from healthcare to personal fitness, gaming, military applications, security fields, etc. HAR can be achieved with sensors, images, smartphones or videos where the advancement of Human Computer Interaction (HCI) technology has become more popular for capturing behaviors using sensors such as accelerometers and gyroscopes. This paper introduces an approach that uses CNN and Long Short-Term Memory (LSTM) to predict human behaviors on the basis of the WISDM dataset.
item: Conference-Full-text
Hybrid approach and architecture to detect fake news on twitter in real-time using neural networks
(Faculty of Information Technology, University of Moratuwa., 2020-12) Thilakarathna, MP; Wijayasekara, VA; Gamage, Y; Peiris, KH; Abeysinghe, C; Rafaideen, I; Vekneswaran, P; Karunananda, AS; Talagala, PD
Fake news has been a key issue since the dawn of social media. Currently, we are at a stage where it is merely impossible to differentiate between real and fake news. This directly and indirectly affects people's decision patterns and makes us question the credibility of the news shared via social media platforms. Twitter is one of the leading social networks in the world by active users. There has been an exponential spread of fake news on Twitter in the recent past. In this paper, we will discuss the implementation of a browser extension which will identify fake news on Twitter using deep learning models with a focus on real-world applicability, architectural stability and scalability of such a solution. Experimental results show that the proposed browser extension has an accuracy of 86% accuracy in fake news detection. To the best of our knowledge, our work is the first of its kind to detect fake news on Twitter real-time using a hybrid approach and evaluate using real users.
item: Conference-Full-text
Identification of brain tumor and extracting its’ features through processing of mri
(Faculty of Information Technology, University of Moratuwa., 2020-12) Lakmi, KWDT; Pathirana, GPSN; Sandanayake, TC; Karunananda, AS; Talagala, PD
The abnormal growth of tissues inside the brain is known as brain tumors and they are considered as a life threatening disease. According to the cell types containing in a tumor they can be classified into two groups as Benign and malignant. Benign tumors are considered to be non-cancerous and they have a primitive shape and size. At the same time malignant tumors are considered to be cancerous and do not have clearly defined edges. Modern technology has introduced several types of imaging techniques for internal body evaluation and analysis. Among them Magnetic Resource Imaging techniques are used to analyze many diseases as they have high resolution and better quality compared to others. Using conventional methods to identify brain tumors using MRI and extracting their features are difficult as the brain is complex. Therefor image processing techniques can be used to detect brain tumors and extract features automatically and effectively. This study presents a method to detect and extract features of the brain tumors which consist of five steps: preprocessing, skull stripping, detecting tumors in axial, coronal and sagittal planes, identifying tumor location and extracting features. The outcomes of the research study will help the doctors or the medical technicians to identify the brain tumor and its features in an effective manner.
item: Conference-Full-text
Infinity yoga tutor: yoga posture detection and correction system
(Faculty of Information Technology, University of Moratuwa., 2020-12) Rishan, F; De Silva, B; Alawathugoda, S; Nijabdeen, S; Rupasinghe, L; Liyanapathirana, C; Karunananda, AS; Talagala, PD
Popularity of yoga is increasing daily. The reason for this is the physical, mental and spiritual benefits that could be obtained by practicing yoga. Many are following this trend and practicing yoga without the training of an expert practitioner. However, following yoga in an improper way or without a proper guidance will lead to bad health issues such as strokes, nerve damage etc. So, following proper yoga postures is an important factor to be considered. In this proposed system, the system is able to identify poses performed by the user and also guide the user visually. This process is required to be completed in real-time in order to be more interactive with the user. In this paper, the yoga posture detection was done in a vision-based approach. The Infinity Yoga Tutor application is able to capture user movements using the mobile camera, which is then streamed at a resolution of 1280 × 720 at 30 frames per second to the detection system. The system consists of two main modules, a pose estimation module which uses OpenPose to identify 25 keypoints in the human body, using the BODY_25 dataset, and a pose detection module which consists of a Deep Learning model, that uses time-distributed Convolutional Neural Networks, Long Short Term Memory and SoftMax regression in order to analyze and predict user pose or asana using a sequence of frames. This module was trained to classify 6 different asanas and the selected model which uses OpenPose for pose estimation has an accuracy of 99.91%. Finally, the system notifies the users on their performance visually in the user interface of the Mobile application.
item: Conference-Full-text
Iot enabled an open framework for air pollution monitoring system
(Faculty of Information Technology, University of Moratuwa., 2020-12) Sudantha, BH; Manchanayaka, MALSK; Premakumara, N; Shiranthika, C; Premachandra, C; Kawanaka, H; Karunananda, AS; Talagala, PD
Blessing with fresh air to breathe is one of the primary living requirements of the Human being. Nowadays, majority of the countries in the world are suffering from the problem of Air pollution. This problem is getting worse day by day because of rapid economic growth, industrialization, urbanization and the resulting rise in energy demand. Air pollution has been become one of the primary concern in Sri Lanka. In most of the Sri Lankan cities, the primary cause for the air pollution is the lack of the prevalence of proper environmental regulations. Although the scientific evaluation of such technologies and systems is generally recognized as a significant facet of open source technologies and their implementation, it is often underexplored in science. This research provides an integrated approach to the development of an Environmental Monitoring System prototype based on open source hardware and software and to track the reliability of the system in terms of data accuracy. This system is able to measure six environmental parameters namely air temperature, air CO Percentage, air NO 2 percentage, air O 3 percentage, air PM percentage and air SO 2 percentage. This research has shown a promising way to create a dense coverage for more cost-effective monitoring of environmental phenomena. Most of the existing monitoring systems for air pollution have inferior accuracy, low sensitivity and require laboratory study. Our implemented system is a three-phase air pollution monitoring system, which have drastically shown improvements over the existing methodologies combining with the Internet of Things (IOT).
item: Conference-Full-text
Iot-enhanced smart laser fence for reducing human elephant conflicts
(Faculty of Information Technology, University of Moratuwa., 2020-12) Firdhous, MFM; Karunananda, AS; Talagala, PD
Human animal conflict is a serious problem in several countries in the Asian and African continents. The human animal conflict commonly results in the damage to properties and loss of human/animal life and limbs. The conflict between elephants and humans is considered to be one of the most serious conflicts due to the nature of losses incurred. The most commonly used deterrent used to keep the animals away from human settlements is the electric fence. Though, electric fences could act as a barrier for elephants to enter into villages, with time elephants learn to break them and enter villages. This paper presents the design of a smart laser fence that can be used for detecting elephants as well as chasing them back to the jungle with the help of other associated systems. The tests carried out on the system show that the concept is working and can be deployed in the field after an extensive field test.