1,721,002 research outputs found
Saliency-based search free car license plate localisation.
No full textAdvances in intelligent transportation systems (ITS) has great societal impact. Automatic license plate recognition (ALPR) or automatic number plate recognition (ANPR) systems have brought valuable insight for the red-light or speed limit enforcement, electronic payment systems, and traffic surveillance. These systems comprise of four separate
phases: image acquisition, localisation, segmentation, and character recognition. Blocked or covered licence plates (LPs) and images taken under bad environmental situations, lightings, size and orientation variations are some of the bottlenecks of ALPR systems. In real-time operations e.g. ITS, ALPR process has to operate rapidly and accurately. Real-time applications of ALPR systems require the fast and accurate detection of LPs becomes even one single miss detection or lack of detection might
cause problems. That is why the localisation phase of ALPR systems seem to be the
most crucial one, determining the speed and the accuracy of the whole system. Popular search-dependent algorithms cannot afford to keep up with the challenges of ALPR systems, due to their computational complexity. This motivates development of search-free methods in this thesis. The proposed methods can be categorised into five groups; algorithms based on: saliency map and local variance; incorporating hierarchical saliency; multiple LPL (license plate localisation) via likelihood estimation; incorporating the differences in distributions, such as by negentropy, and finally through a deep convolution neural network approach for LPL. The first proposed algorithm does not need to search the image and it is based on estimations of saliency and local variance (using L1 − norm). It also uses Gabor
function to exploit directionality for validating and choosing the right LP. In this algorithm, traditional saliency map detection (SMD) method locates the cars in images in a very fast, search-free approach. Next proposed approach modifies the traditional SMD method by means of directionality and a new definition of saliency, called saliency based ALPR (SB-ALPR). In addition, L1 −norm is used to better identify the pattern of characters and numbers in the LP regions. Experimental results show an average 94.77% accuracy, 61.52 ms execution time and 40.2 ms processing time per LP, which is a remarkable achievement in comparison to the performance of the previous algorithm using traditional SMD method. A search-free LP localisation algorithm on the basis of 3-D Bayesian saliency estimation is the next proposed method. In this algorithm, 3-D objects are traced by means of object/shadow detection and removal and Bayesian method for object recognition. This algorithm consists of three fundamental phases: object/shadow detection and object recognition. For the object detection phase, the background image and the moving objects are detected. To eliminate the shadows, a new approach,
in which we discriminate the shadows from their corresponding objects, is exploited and for tracking purpose, the relations between the objects in sequential frames are determined via Bayesian method. The results show that, unlike the previous algorithm, in this proposed method the image backgrounds are more accurately subtracted and
the shadows eliminated by an accuracy of approximately 70%. Object recognition phase is also subjoined to the algorithm, which performed well. In another novel approach the distance between the distributions by means of negentropy is combined with the saliency based ALPR. We then utilise the Bayesian
decision making method for selecting the LP from the detected candidates. This proposed algorithm shows an accuracy of 96% and a computation time of 80ms per plate, which is an outstanding improvement over the classic touchstone techniques. In the last proposed method a deep convolution neural network is successfully used for LPL. Following this approach the network learns the necessary filters at each layer and generates new patterns by convolving the filters with the images in previous layers
Eigen-based machine learning techniques for complex and hyper-complex processing.
Full text linkOne of the earlier works on eigen-based techniques for the hyper-complex domain of quaternions was on “quaternion principal component analysis of colour images”. The results of this work are still instructive in many aspects. First, it showed how naturally the quaternion domain accounts for the coupling between the dimensions of red, blue and green of an image, hence its suitability for multichannel processing. Second, it was clear that there was a lack of eigen-based techniques for such a domain, which explains the non-trivial gap in the literature. Third, the lack of such eigen-based quaternion tools meant that the scope and
the applications of quaternion signal processing were quite limited, especially in the field of biomedicine. And fourth, quaternion principal component analysis made use of complex matrix algebra, which reminds us that the complex domain lays the building blocks of the quaternion domain, and therefore any research endeavour in quaternion signal processing should start with the complex domain.
As such, the first contribution of this thesis lies in the proposition of complex singular spectrum analysis. That research provided a deep understanding and an appreciation of the intricacies of the complex domain and its impact on the quaternion domain. As the complex domain offers one degree of freedom over the real domain, the statistics of a complex variable x has to be augmented with its complex conjugate x*, which led to the term augmented statistics. This recent advancement in complex statistics was exploited in the proposed complex singular spectrum analysis. The same statistical notion was used in proposing novel quaternion eigen-based techniques such as the quaternion singular spectrum analysis, the quaternion uncorrelating transform, and the quaternion common spatial patterns. The latter two methods highlighted an important gap in the literature – there were no algebraic methods that solved the simultaneous diagonalisation of quaternion matrices. To address this issue, this thesis also presents new fundamental results on quaternion matrix factorisations and explores the depth of quaternion algebra.
To demonstrate the efficacy of these methods, real-world problems mainly in biomedical engineering were considered. First, the proposed complex singular spectrum analysis successfully addressed an examination of schizophrenic data
through the estimation of the event-related potential of P300. Second, the automated detection of the different stages of sleep was made possible using the proposed quaternion singular spectrum analysis. Third, the proposed quaternion common spatial patterns facilitated the discrimination of Parkinsonian patients from healthy subjects. To illustrate the breadth of the proposed eigen-based techniques, other areas of applications were also presented, such as in wind and financial forecasting, and Alamouti-based communication problems. Finally, a
preliminary work is made available to suggest that the next step from this thesis is to move from static models (eigen-based models) to dynamic models (such as tracking models)
Blind source separation via independent and sparse component analysis with application to temporomandibular disorder
Full text linkBlind source separation (BSS) addresses the problem of separating multi channel signals observed by generally spatially separated sensors into their constituent underlying sources. The passage of these sources through an unknown mixing medium results in these observed multichannel signals. This study focuses on BSS, with special emphasis on its application to the temporomandibular joint disorder (TMD). TMD refers to all medical problems related to the temporomandibular joint (TMJ), which holds the lower jaw (mandible) and the temporal bone (skull). The overall objective of the work is to extract the two TMJ sound sources generated by the two TMJs, from the bilateral recordings obtained from the auditory canals, so as to aid the clinician in diagnosis and planning treatment policies. Firstly, the concept of 'variable tap length' is adopted in convolutive blind source separation. This relatively new concept has attracted attention in the field of adaptive signal processing, notably the least mean square (LMS) algorithm, but has not yet been introduced in the context of blind signal separation. The flexibility of the tap length of the proposed approach allows for the optimum tap length to be found, thereby mitigating computational complexity or catering for fractional delays arising in source separation. Secondly, a novel fixed point BSS algorithm based on Ferrante's affine transformation is proposed. Ferrante's affine transformation provides the freedom to select the eigenvalues of the Jacobian matrix of the fixed point function and thereby improves the convergence properties of the fixed point iteration. Simulation studies demonstrate the improved convergence of the proposed approach compared to the well-known fixed point FastICA algorithm. Thirdly, the underdetermined blind source separation problem using a filtering approach is addressed. An extension of the FastICA algorithm is devised which exploits the disparity in the kurtoses of the underlying sources to estimate the mixing matrix and thereafter achieves source recovery by employing the i-norm algorithm. Additionally, it will be shown that FastICA can also be utilised to extract the sources. Furthermore, it is illustrated how this scenario is particularly suitable for the separation of TMJ sounds. Finally, estimation of fractional delays between the mixtures of the TMJ sources is proposed as a means for TMJ separation. The estimation of fractional delays is shown to simplify the source separation to a case of in stantaneous BSS. Then, the estimated delay allows for an alignment of the TMJ mixtures, thereby overcoming a spacing constraint imposed by a well- known BSS technique, notably the DUET algorithm. The delay found from the TMJ bilateral recordings corroborates with the range reported in the literature. Furthermore, TMJ source localisation is also addressed as an aid to the dental specialist.EThOS - Electronic Theses Online ServiceGBUnited Kingdo
Blind source separation via independent and sparse component analysis with application to temporomandibular disorder
Full text linkBlind source separation (BSS) addresses the problem of separating multi channel signals observed by generally spatially separated sensors into their constituent underlying sources. The passage of these sources through an unknown mixing medium results in these observed multichannel signals. This study focuses on BSS, with special emphasis on its application to the temporomandibular joint disorder (TMD). TMD refers to all medical problems related to the temporomandibular joint (TMJ), which holds the lower jaw (mandible) and the temporal bone (skull). The overall objective of the work is to extract the two TMJ sound sources generated by the two TMJs, from the bilateral recordings obtained from the auditory canals, so as to aid the clinician in diagnosis and planning treatment policies. Firstly, the concept of 'variable tap length' is adopted in convolutive blind source separation. This relatively new concept has attracted attention in the field of adaptive signal processing, notably the least mean square (LMS) algorithm, but has not yet been introduced in the context of blind signal separation. The flexibility of the tap length of the proposed approach allows for the optimum tap length to be found, thereby mitigating computational complexity or catering for fractional delays arising in source separation. Secondly, a novel fixed point BSS algorithm based on Ferrante's affine transformation is proposed. Ferrante's affine transformation provides the freedom to select the eigenvalues of the Jacobian matrix of the fixed point function and thereby improves the convergence properties of the fixed point iteration. Simulation studies demonstrate the improved convergence of the proposed approach compared to the well-known fixed point FastICA algorithm. Thirdly, the underdetermined blind source separation problem using a filtering approach is addressed. An extension of the FastICA algorithm is devised which exploits the disparity in the kurtoses of the underlying sources to estimate the mixing matrix and thereafter achieves source recovery by employing the i-norm algorithm. Additionally, it will be shown that FastICA can also be utilised to extract the sources. Furthermore, it is illustrated how this scenario is particularly suitable for the separation of TMJ sounds. Finally, estimation of fractional delays between the mixtures of the TMJ sources is proposed as a means for TMJ separation. The estimation of fractional delays is shown to simplify the source separation to a case of in stantaneous BSS. Then, the estimated delay allows for an alignment of the TMJ mixtures, thereby overcoming a spacing constraint imposed by a well- known BSS technique, notably the DUET algorithm. The delay found from the TMJ bilateral recordings corroborates with the range reported in the literature. Furthermore, TMJ source localisation is also addressed as an aid to the dental specialist
Rebuttal to "Comments on The Quaternion LMS Algorithm for Adaptive Filtering of Hypercomplex Processes"
Full text linkThis is the authors' rebuttal on the correspondence article "Comments on The Quaternion LMS Algorithm for Adaptive Filtering of Hypercomplex Processes" by Gang Wang and Rui Xue
Simultaneous diagonalisation of the covariance and complementary covariance matrices in quaternion widely linear signal processing
Full text linkEnergy and performance aware resource management in heterogeneous cloud datacenters.
Full text linkIn cloud computing, datacenters are the principal consumers of electricity. In 2014, Cloud datacenters reportedly accounted for some 70 billion kWh, which is the equivalent of 1.8% of the US’ total energy consumption. With growth in on-line services, but increased computational power per unit of energy, consumption is projected to account for 73 billion kWh by 2020. Datacenters comprise large numbers of servers, as well as storage, that cloud customers can use in the amounts they require for as long as they are willing to pay. In infrastructure clouds, customers request the launch of Virtual Machines (VMs) which will consume server and storage resources. The provider decides which server is selected, and the customer decides how long to run the VM for. The unpredictability of customers of infrastructure clouds can result in datacenters having a number of servers either idle or running a minimal VM loading at various times, and wasting energy as a consequence. Improvements to management techniques such as VM allocation and resource consolidation can help to improve energy and performance efficiency. However, for a particular VM the energy consumption and runtime may be different in different servers due to: (i) the number of VMs the servers run; and (ii) the performance of servers. Therefore, w.r.t VM allocation it might be more energy and performance efficient to place VMs on servers that consume less energy and can meet the VM performance goals. Moreover, consolidation brings two, related, problems: (i) consolidation involves migrating VMs across servers, which adds to energy consumption, and will only be more energy efficient if this cost can be recovered; and (ii) due to resource heterogeneity the performance of VMs varies with the underlying hardware, and with it, runtimes and energy usage, and hence costs. In respect to (i), if the VM terminates during or just after the migration has finished, the migration effort is definitely wasted, which implies a cost recovery time objective after which further energy can be saved as the VM subsequently runs more efficiently. In respect to (ii), if the VM is migrated to a server with lower performance, increased runtime can decrease datacenter throughput and energy efficiency, and increase agreed (pay per use) customer cost. We explore how consolidation of VMs can help to decrease datacenter energy consumption whilst ensuring that migration costs are recoverable in the vast majority of cases, and also ensuring that workload performance is not negatively affected. Several algorithms for energy-performance efficient VM allocation and consolidation are proposed, implemented through extensions and modifications to the popular Cloud simulation environment, CloudSim, and evaluated in respect to a large dataset of workload information from a major cloud provider. Principal findings from these simulations are: (i) efficient VM allocation can be at least 1.72% (±0.02 error) more energy-efficient than consolidation; (ii) it is 3.52% (±0.05 error) more energy-efficient to migrate relatively long-running VMs; and (iii) for heterogeneous workloads and clouds, different scheduling and migration techniques demonstrate a diversity in energy efficiency and performance (hence cost) trade-off. An energy-performance efficient migration approach can be up to 3.66% (±0.05 error) more energy efficient, and 1.87% (±0.025 error) more performance efficient, than a no migration strategy. This suggests a saving of approximately 1.58m/year. Based on these results, cloud providers could both reduce their energy usage, reducing costs and either pass savings to customers, invest in more infrastructure, or increase profits; more broadly, such reductions in energy usage could reduce the impact of global warming
Widrow-Hoff LMS Adaline Demonstrator for Schools and Colleges
Full text linkThe Widrow-Hoff LMS (or ‘Adaline’) algorithm developedoriginally in 1960 is fundamental to the operation of countlesssignal processing machine learning systems in use eventoday. Bernard Widrow and Ted Hoff famously developedan Adaline machine demonstrator using basic analog offthe shelf components to show how a ‘perceptron’ couldbe trained manually [1]. This paper details the design anddevelopment of a fully digital Adaline Least-Mean-Squarealgorithm demonstrator. The simplistic design presented hereis completely open-source with all code, bill of materialsand 3D casing models available at this Github Repository fordownload and reproduction. The demonstrator enables quickvisualisation of the training and testing of a simple perceptronalgorithm running on the inexpensive Arduino platform. Thetotal costs of the device is estimated to be less than $60 andcould be used in classrooms and colleges the world-over todemonstrate the seminal work of Widrow and Hoff [2] to awide audience
Widrow-Hoff LMS Adaline Demonstrator for Schools and Colleges
Full text linkThe Widrow-Hoff LMS (or ‘Adaline’) algorithm developedoriginally in 1960 is fundamental to the operation of countlesssignal processing machine learning systems in use eventoday. Bernard Widrow and Ted Hoff famously developedan Adaline machine demonstrator using basic analog offthe shelf components to show how a ‘perceptron’ couldbe trained manually [1]. This paper details the design anddevelopment of a fully digital Adaline Least-Mean-Squarealgorithm demonstrator. The simplistic design presented hereis completely open-source with all code, bill of materialsand 3D casing models available at this Github Repository fordownload and reproduction. The demonstrator enables quickvisualisation of the training and testing of a simple perceptronalgorithm running on the inexpensive Arduino platform. Thetotal costs of the device is estimated to be less than $60 andcould be used in classrooms and colleges the world-over todemonstrate the seminal work of Widrow and Hoff [2] to awide audience
Extracting information from heterogeneous internet of things data streams.
Full text linkRecent advancements in sensing, networking technologies and collecting real-world data on a large scale and from various environments have created an opportunity for new forms of services and applications. This is known under the umbrella term of the Internet of Things (IoT). Physical sensor devices constantly produce very large amounts of data. Methods are needed which give the raw sensor measurements a meaningful interpretation for building automated decision support systems. One of the main research challenges in this domain is to extract actionable information from real-world data, that is information that can readily be used to make informed automatic decisions in intelligent systems. Most existing approaches are application or domain dependent or are only able to deal with specific data sources of one kind. This PhD research concerns multiple approaches for analysing IoT data streams. We propose a method which determines how many different clusters can be found in a stream based on the data distribution. After selecting the number of clusters, we use an online clustering mechanism to cluster the incoming data from the streams. Our approach remains adaptive to drifts by adjusting itself as the data changes. The work is benchmarked against state-of-the art stream clustering algorithms on data streams with data drift. We show how our method can be applied in a use case scenario involving near real-time traffic data. Our results allow to cluster, label and interpret IoT data streams dynamically according to the data distribution. This enables to adaptively process large volumes of dynamic data online based on the current situation. We show how our method adapts itself to the changes and we demonstrate how the number of clusters in a real-world data stream can be determined by analysing the data distributions. Using the ideas and concepts of this approach as a starting point we designed another novel dynamic and adaptable clustering approach that is more suitable for multi-variate time-series data clustering. Our solution uses probability distributions and analytical methods to adjust the centroids as the data and feature distributions change over time. We have evaluated our work against some well-known time-series clustering methods and have shown how the proposed method can reduce the complexity and perform efficient in multi-variate datastreams. Finally we propose a method that uncovers hidden structures and relations between multiple IoT data streams. Our novel solution uses Latent Dirichlet Allocation (LDA), a topic extraction method that is generally used in text analysis. We apply LDA on meaningful labels that describe the numerical data in human understandable terms. To create the labels we use Symbolic Aggregate approXimation (SAX), a method that converts raw data into string-based patterns. The extracted patterns are then transformed with a rule engine into the labels. The work investigates how heterogeneous sensory data from multiple sources can be processed and analysed to create near real-time intelligence and how our proposed method provides an efficient way to interpret patterns in the data streams. The proposed method provides a novel way to uncover the correlations and associations between different pattern in IoT data streams. The evaluation results show that the proposed solution is able to identify the correlation with high efficiency with an F-measure up to 90%. Overall, this PhD research has designed, implemented and evaluated unsupervised adaptive algorithms to analyse, structure and extract information from dynamic and multi-variate sensory data streams. The results of this research has significant impact in designing flexible and scalable solutions in analysing real-world sensory data streams and specially in cases where labelled and annotated data is not available or it is too costly to be collected. Research and advancements in healthcare and smarter cities are two key areas that can directly from this research
- …
