Department Publications

In this paper, we introduce a novel algorithm that can dramatically reduce the number of antenna elements needed to accurately predict the direction of arrival (DOA) for multiple input multiple output (MIMO) radar. The new proposed algorithm predicts the received signal of a large antenna setup using reduced number of antenna by using coupled dictionary learning. Hence, this enables the MIMO radar to resolve more paths, which could not be resolved by the fewer antennas. Specifically, we overcome the problem of inaccurate DOA estimation due to a small virtual array setup. For example, we can use dictionary learning to predict 100 virtual array elements using only 25. To evaluate our algorithm, we used multiple signal classification (MUSIC) as a DOA estimation technique to estimate the DOA for non coherent multiple targets. The results show that using the predicted received signal, the proposed algorithm …

Machine vision has been widely deployed in many industrial applications. However, the accuracy and maturity level for the inspection of through-hole technology (THT) solder joints have yet to reach its ultimate goal. In this paper, we have presented a detailed explanation on a set of novel algorithms that can be effectively used to implement an automatic vision system that is capable of classifying the quality of THT solder joints very precisely. This vision system can be easily integrated into a soldering robotic system that performs automatic soldering on THT solder joints. The proposed vision system is capable of controlling the entire robotic system while providing automatic rework capability on the defective solder joints with much less user interaction. The experimental results indicate that the proposed system has an improved recognition rate and good resilience to noise and uneven illumination distribution.

In recent years, Convolution Neural Network (CNN) gained great success in many applications, especially in computer vision. Now adapting CNN inference on edge devices has become an active research area in embedded vision and hot topic in Edge AI. The primary design hurdles for implementing CNN inference on embedded systems are limited computational resource, memory resource, and power budget. This study presents a novel architecture for SqueezeNet [1] like CNN models, which can be extended to support any CNN model as well. We address two approaches to mitigate resource constraints. First, we use a custom floating point(12 bit for computation and 8bit for storing). Second is slicing the model into a repetitive block called computation blocks. Computation block can be configured dynamically by the host processor to operate in a different mode. We have implemented SqueezeNet v1.1 for …

This paper presents the development of a Sinhala Speech Recognition System to be deployed in an Interactive Voice Response (IVR) system of a telecommunication service provider. The main objectives are to recognize Sinhala digits and names of Sinhala songs to be set up as ringback tones. Sinhala being a phonetic language, its features are studied to develop a list of 47 phonemes. A continuous speech recognition system is developed based on Hidden Markov Model (HMM). The acoustic model is trained using the voice through mobile phone. The outcome is a speaker independent speech recognition system which is capable of recognizing 10 digits and 50 Sinhala songs. A word error rate (WER) of 11.2% using a speech corpus of 0.862 hours and a sentence error rate (SER) of 5.7% using a speech corpus of 1.388 hours are achieved for digits and songs respectively.

Purpose Automatic Optical Inspection (AOI) systems that are extensively used in the industry of Electronics Manufacturing Services (EMS), performs the inspection of Surface Mount Devices (SMD). One of the main tasks of such an AOI system is to align a given PCB to the parameters of the corresponding PCB positioning system by a process called fiducial alignment. However, no detailed analysis has been carried out so far on the methodologies that can be used to have a very precise identification of PCB fiducial points. In our research, we have implemented an AOI system for the inspection of soldering defects of Through Hole Technology (THT) solder joints, which can be integrated to a desktop soldering robotic platform. Such platforms are used in environments where no specific lighting conditions can be provided within a surrounded atmosphere. Therefore, an AOI system that is capable of performing fiducial alignment of any given PCB under varying lighting condition is highly preferred. In this paper, we have presented a detailed analysis on feature extraction and template matching algorithms that can be used to implement a very precise fiducial verification process under normal lighting condition. Design/methodology/approach A detailed analysis and performance evaluation have been carried out in this paper on prominent image comparison algorithms that are extensively used in the field of image processing. Findings According to the analysis carried out in this paper, it could be observed that the combination of feature extraction and template matching algorithms gives the best performance in PCB fiducial verification process. Research …

Convolutional Neural Networks (CNNs) are widely used in deep learning applications, e.g. visual systems, robotics etc. However, existing software solutions are not efficient. Therefore, many hardware accelerators have been proposed optimizing performance, power and resource utilization of the implementation. Amongst existing solutions, Field Programmable Gate Array (FPGA) based architecture provides better cost-energy-performance trade-offs as well as scalability and minimizing development time. In this paper, we present a model-independent reconfigurable co-processing architecture to accelerate CNNs. Our architecture consists of parallel Multiply and Accumulate (MAC) units with caching techniques and interconnection networks to exploit maximum data parallelism. In contrast to existing solutions, we introduce limited precision 32 bit Q-format fixed point quantization for arithmetic representations and …

Biological sampling for biomass estimation is a highly labor intensive process that can benefit from automation using computer vision. In this paper two algorithms the Green Layer Extraction (GLE) method and Morphology and Marker Controlled Watershed Transformation (MMCWT) method are proposed for the estimation of frond (leaf) area of an invasive aquatic fern Salvinia molesta. The two algorithms are tested using a computer controlled automatic testbed. The mean accuracy for the GLE was found to be 89.66% for estimating the number of fronds and 86.74 % for estimating the frond area. For the MMCWT the respective results are 91.92% and 80.68%. Then the results are subsequently used for developing a mathematical model as growth rate estimation of Salvinia molesta for various concentrations of Sodium di-hydrogen phosphate.

Activity recognition in videos in a deep-learning setting-or otherwise-uses both static and pre-computed motion components. The method of combining the two components, while keeping the burden on the deep network less, still remains uninvestigated. Moreover, it is not clear what the level of contribution of individual components is, and how to control the contribution. In this paper, we use a combination of convolutional-neural-network-generated static features and motion features in the form of motion tubes. We propose three schemas for combining static and motion components: based on a variance ratio, principal components, and Cholesky decomposition. The Cholesky-decomposition-based method allows the control of contributions. The ratio given by variance analysis of static and motion features matches well with the experimental optimal ratio used in the Cholesky decomposition-based method. The …

Purpose Automatic Optical Inspection (AOI) systems, used in electronics industry have been primarily developed to inspect soldering defects of Surface Mount Devices (SMD) on a Printed Circuit Board (PCB). However, no commercially available AOI system exists that can be integrated to a desktop soldering robotic system, which is capable of identifying soldering defects of Through Hole Technology (THT) solder joints along with the soldering process. In our research, we have implemented an AOI platform that is capable of performing automatic quality assurance of THT solder joints in a much efficient way. In this paper, we have presented a novel approach to identify soldering defects of THT solder joints, based on the location of THT component lead top. This paper presents the methodologies that can be used to precisely identify and localize THT component lead inside a solder joint. Design/methodology/approach We have discussed the importance of lead top localization and presented a detailed description on the methodologies that can be used to precisely segment and localize THT lead top inside the solder joint. Findings It could be observed that the precise localization of THT lead top makes the soldering quality assurance process more accurate. A combination of template matching algorithms and colour model transformation provide the most accurate outcome in localizing the component lead top inside solder joint, according to the analysis carried out in this paper. Research limitations/implications When the component lead top is fully covered by the soldering, the implemented methodologies will not be able to identify the actual …

Many real world stochastic processes are non-stationary, which means that the probability distribution that generates data samples is time-varying. In the context of machine learning, this phenomenon is known as concept drift. It is important that machine learning models are able to adapt to concept drift in order to prevent degradation in accuracy. In this paper, we present two algorithms for drift detection and adaptation. Drift is measured by continuously tracking a difference metric between probability distributions estimated from two sample windows preceding a time point. High values for the difference metric indicates that concept drift has occurred, and the model must be adapted. Adaptation is done by training a new model for the drifted process, and adding it to an ensemble of models. Previously trained models are retained, and their weights in the ensemble are adjusted to reflect similarity with the current …

A complex activity is a temporal composition of sub-events, and a sub-event typically consists of several low level micro-actions, such as body movement of different actors. Extracting these micro actions explicitly is beneficial for complex activity recognition due to actor selectivity, higher discriminative power, and motion clutter suppression. Moreover, considering both static and motion features is vital for activity recognition. However, optimally controlling the contribution from static and motion features still remains uninvestigated. In this work, we extract motion features at micro level, preserving the actor identity, to later obtain a high-level motion descriptor using a probabilistic model. Furthermore, we propose two novel schemas for combining static and motion features: Cholesky-transformation based and entropy-based. The former allows to control the contribution ratio precisely, while the latter obtains the optimal …

The development of computer based systems expands the usage of computer based application in human life. It can be observed that illegal activities such as unauthorized data access, data theft, data modification and various other intrusion activities are rapidly growing during last decade. Hence, deployment and continuous improvement of Intrusion Detection Systems (IDS) are of paramount importance. Training, testing and evaluation of IDS with real network traffic is significant challenge, so most of IDS evaluation is based on intrusion datasets. Therefore, analysis of intrusion datasets are of paramount importance. In this paper, we evaluated Aegean Wi-Fi Intrusion Dataset (AWID) with different machine learning techniques. Feature reduction techniques such as Information Gain (IG) and Chi-Squared statistics (CH) were applied to evaluate dataset performance with feature reduction. Results of experiments show …

The P300 wave is an Event Related Potential that is elicited in the brain when a subject is presented with a familiar stimulus. The Guilty Knowledge Test is used to determine if certain information is stored in the brain by detecting the P300 wave. In this paper, an improved Guilty Knowledge Test has been conducted on 14 subjects. Although a variety of feature extraction and classification algorithms for detecting the P300 have been used thus far, this paper introduces the use of deep learning techniques for the Guilty Knowledge Test. Two deep learning techniques; Deep Belief Networks and Stacked Autoencoders, have been compared. The input to these classifiers was the filtered EEG signal without any feature extraction. A mean accuracy of 86.9% was achieved for the Deep Belief Network and 86.01% for the Stacked Autoencoder on our dataset, outperforming the conventional Support Vector Machine.

Intrusion detection is only a starting step in securing IT infrastructure. Prediction of intrusions is the next step to provide an active defense against incoming attacks. Current intrusion prediction methods focus mainly on prediction of either intrusion type or intrusion category and do not use or provide contextual information such as source and target IP address. In addition most of them are dependant on domain knowledge and specific scenario knowledge. The proposed algorithm employs a bag-of-words model together with a hidden Markov model which not depend on specific domain knowledge. Since this algorithm depends on a training process it is adaptable to different conditions. A key advantage of the proposed algorithm is the inclusion of contextual data such as source IP address, destination IP range, alert type and alert category in its prediction, which is crucial for an eventual response. Experiments conducted using a public data set generated over 2500 alert predictions and achieved accuracy of 81% and 77% for single step and five step predictions respectively for prediction of the next alert cluster. It also achieved an accuracy of prediction of 95% and 92% for single step and five step predictions respectively for prediction of the next alert category. The proposed methods achieved a prediction accuracy improvement of 5% for alert category over existing variable length Markov chain intrusion prediction methods, while providing more information for a possible defense.

Intrusion detection is only a starting step in securing IT infrastructure. Prediction of intrusions is the next step to provide an active defense against incoming attacks. Most of the existing intrusion prediction methods mainly focus on prediction of either intrusion type or intrusion category. Also, most of them are built based on domain knowledge and specific scenario knowledge. This thesis proposes an alert prediction framework which provides more detailed information than just the intrusion type or category to initiate possible defensive measures. The proposed algorithm is based on hidden Markov model and it does not depend on specific domain knowledge. Instead, it depends on a training process. Hence the proposed algorithm is adaptable to different conditions. Also, it is based on prediction of the next alert cluster, which contains source IP address, destination IP range, alert type and alert category. Hence …

The P300 Speller is a Brain Computer Interface that enables communication using the EEG signal. The P300 wave is an Event Related Potential that occurs as a response to a familiar stimulus. This system can be used to aid persons who are unable to communicate via conventional methods. In this paper, the P300 Speller has been modified to allow communication in three languages: English, Sinhala and Tamil. Classification of the P300 waveform was done using an ensemble of Support Vector Machines. A mean letter classification accuracy of 83.76% and a mean P300 classification accuracy of 90.68% was achieved on our dataset.

The rapid advancement of technology has made the telecommunication sector very competitive. In order to keep up with the competition, telecommunication operators have to identify the exact needs of the customers and offer services in-line with the customer needs. The aim of this research is to find out how data mining can be applied to identify customer needs and how telecommunication companies can benefit from using data mining techniques in their business and how to develop a data mining framework for identifying the customer needs in order to develop products and services to increase the revenue levels of the companies. The objective of this study includes the investigation of the relationship between data mining practices and the customer behavior patterns, the relationship between customer needs and products or services, relationship between new product design initiatives and revenue increases of the companies, the impact of data mining on the revenue of telecommunications companies and the development of the data mining frame work to improve the overall ARPU levels of the industry and a Business Intelligence (BI) tool to enhance decision making for improving the overall ARPU levels of the industry. Firstly the conceptual model is developed to indicate the main information areas that the study seeks to provide insights on and is tested through a sample of employees who hold positions in the telecommunications sector. This model has four main variables; an independent variable (data mining), two intermediate variables (customer behavior and product) and a dependent variable (increased revenue). Secondly the …

We investigate the problem of automatic action recognition and classification of videos. In this paper, we present a convolutional neural network architecture, which takes both motion and static information as inputs in a single stream. We show that the network is able to treat motion and static information as different feature maps and extract features off them, although stacked together. We trained and tested our network on Youtube dataset. Our network is able to surpass state-of-the-art hand-engineered feature methods. Furthermore, we also studied and compared the effect of providing static information to the network, in the task of action recognition. Our results justify the use of optic flows as the raw information of motion and also show the importance of static information, in the context of action recognition.

Automatic stroke recognition of badminton video footages plays an important role in the process of analyzing players and building up statistics. Yet recognizing activities from broadcast videos is a challenging task due to person dependant body postures and blurring of the fast moving body parts. We propose a robust and an accurate approach for badminton stroke recognition using dense trajectories and trajectory aligned HOG features which are calculated inside local bounding boxes around players. A four-class SVM classifier is then used to classify badminton strokes to be either smash, forehand, backhand or other. This approach is robust to noisy backgrounds and provides accurate results for low resolution broadcast videos. Our experiments also reveal that this approach needs relatively fewer training samples for accurate recognition of strokes compared to existing approaches.

Abnormal activity detection plays an important role in many areas such as surveillance, military installations, and sports. Existing abnormal activity detectors mostly rely on motion data obtained over a number of frames to characterize abnormality. However, only motion may not be able to capture all forms of abnormality, in particular, poses that do not amount to motion "outliers". In this paper, we propose two different spatio-temporal descriptors, a silhouette and optic flow based method and a dense trajectory based method which additionally include trajectory shape descriptor, to detect abnormalities. These two descriptors enable us to classify abnormal versus non-abnormal activities using SVM. Comparison with existing methods, using five standard datasets, shows that dense trajectory based method outperforms state-of-the-art results in crowd dataset and silhouette and optic flow based method outperforms …

Transferring computational tasks from user-worn devices to everyday objects allows users to freely focus on their regular non-computing tasks. Identifying micro-activities (short-repetitive-activities that compose the macro-level behavior) enables the understanding of subtle behavioral changes and providing just-in-time information without explicit user input. In this paper, we propose the concept of micro-activity recognition of augmented-everyday-objects and evaluate the applicability of machine learning algorithms that are previously used for macro-level activity recognition. We outline a few proof-of-concept application scenarios that provide micro-activity-aware just-in-time information.

Action recognition in a video plays an important role in computer vision and finds many applications in areas such as surveillance, sports, and elderly monitoring. Existing methods mostly rely on stationary backgrounds. Action recognition in dynamic backgrounds typically requires standard preprocessing steps such as motion compensation, background modeling, moving object detection and object recognition. The errors of the motion compensation step and background modelling increase the mis-detections. Therefore action recognition in dynamic background is challenging. In this paper, we use a combination of pose characterized by a silhouette and optic flows synthesized into a histogram. This enables us to classify the movement of the actor versus movement of the background. We use four background models to extract the silhouette from the frame. We use SVM to recognize actions, according to several …