Deep learning applied to the semantic segmentation of tyre stockpiles
- Authors: Barfknecht, Nicholas Christopher
- Date: 2018
- Subjects: Neural networks (Computer science)
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: http://hdl.handle.net/10948/23947 , vital:30647
- Description: The global push for manufacturing which is environmentally sustainable has disrupted standard methods of waste tyre disposal. This push is further intensified by the health and safety risks discarded tyres pose to the surrounding population. Waste tyre recycling initiatives in South Africa are on the increase; however, there is still a growing number of undocumented tyre stockpiles developing throughout the country. The plans put in place to eradicate these tyre stockpiles have been met with collection, transport and storage logistical issues caused by the remoteness and distant locales. Eastwood (2016) aimed at optimising the logistics associated with collection, by estimating the number of visible tyres from images of tyre stockpiles. This research was limited by the need for manual segmentation of each tyre stockpile located within each image. This research proposes the use of semantic segmentation to automatically segment images of tyre stockpiles. An initial review of neural network, convolutional network and semantic segmentation literature resulted in the selection of Dilated Net as the semantic segmentation architecture for this research. Dilated Net builds upon the VGG-16 classification architecture to perform semantic segmentation. This resulted in classification experiments which were evaluated using precision, recall and f1-score. The results indicated that regardless of tyre stockpile image dimension, fairly accurate levels of classification accuracy can be attained. This was followed by semantic segmentation experiments which made use of intersection over union (IoU) and pixel accuracy to evaluate the effectiveness of Dilated Net on images of tyre stockpiles. The results indicated that accurate tyre stockpile segmentation regions can be obtained and that the trained model generalises well to unseen images.
- Full Text:
- Date Issued: 2018
A hybridisation technique for game playing using the upper confidence for trees algorithm with artificial neural networks
- Authors: Burger, Clayton
- Date: 2014
- Subjects: Neural networks (Computer science) , Computer algorithms
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: http://hdl.handle.net/10948/3957 , vital:20495
- Description: In the domain of strategic game playing, the use of statistical techniques such as the Upper Confidence for Trees (UCT) algorithm, has become the norm as they offer many benefits over classical algorithms. These benefits include requiring no game-specific strategic knowledge and time-scalable performance. UCT does not incorporate any strategic information specific to the game considered, but instead uses repeated sampling to effectively brute-force search through the game tree or search space. The lack of game-specific knowledge in UCT is thus both a benefit but also a strategic disadvantage. Pattern recognition techniques, specifically Neural Networks (NN), were identified as a means of addressing the lack of game-specific knowledge in UCT. Through a novel hybridisation technique which combines UCT and trained NNs for pruning, the UCTNN algorithm was derived. The NN component of UCT-NN was trained using a UCT self-play scheme to generate game-specific knowledge without the need to construct and manage game databases for training purposes. The UCT-NN algorithm is outlined for pruning in the game of Go-Moku as a candidate case-study for this research. The UCT-NN algorithm contained three major parameters which emerged from the UCT algorithm, the use of NNs and the pruning schemes considered. Suitable methods for finding candidate values for these three parameters were outlined and applied to the game of Go-Moku on a 5 by 5 board. An empirical investigation of the playing performance of UCT-NN was conducted in comparison to UCT through three benchmarks. The benchmarks comprise a common randomly moving opponent, a common UCTmax player which is given a large amount of playing time, and a pair-wise tournament between UCT-NN and UCT. The results of the performance evaluation for 5 by 5 Go-Moku were promising, which prompted an evaluation of a larger 9 by 9 Go-Moku board. The results of both evaluations indicate that the time allocated to the UCT-NN algorithm directly affects its performance when compared to UCT. The UCT-NN algorithm generally performs better than UCT in games with very limited time-constraints in all benchmarks considered except when playing against a randomly moving player in 9 by 9 Go-Moku. In real-time and near-real-time Go-Moku games, UCT-NN provides statistically significant improvements compared to UCT. The findings of this research contribute to the realisation of applying game-specific knowledge to the UCT algorithm.
- Full Text:
- Date Issued: 2014
A feasibility study into total electron content prediction using neural networks
- Authors: Habarulema, John Bosco
- Date: 2008
- Subjects: Electrons , Neural networks (Computer science) , Global Positioning System , Ionosphere , Ionospheric electron density
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5466 , http://hdl.handle.net/10962/d1005251 , Electrons , Neural networks (Computer science) , Global Positioning System , Ionosphere , Ionospheric electron density
- Description: Global Positioning System (GPS) networks provide an opportunity to study the dynamics and continuous changes in the ionosphere by supplementing ionospheric measurements which are usually obtained by various techniques such as ionosondes, incoherent scatter radars and satellites. Total electron content (TEC) is one of the physical quantities that can be derived from GPS data, and provides an indication of ionospheric variability. This thesis presents a feasibility study for the development of a Neural Network (NN) based model for the prediction of South African GPS derived TEC. The South African GPS receiver network is operated and maintained by the Chief Directorate Surveys and Mapping (CDSM) in Cape Town, South Africa. Three South African locations were identified and used in the development of an input space and NN architecture for the model. The input space includes the day number (seasonal variation), hour (diurnal variation), sunspot number (measure of the solar activity), and magnetic index(measure of the magnetic activity). An attempt to study the effects of solar wind on TEC variability was carried out using the Advanced Composition Explorer (ACE) data and it is recommended that more study be done using low altitude satellite data. An analysis was done by comparing predicted NN TEC with TEC values from the IRI2001 version of the International Reference Ionosphere (IRI), validating GPS TEC with ionosonde TEC (ITEC) and assessing the performance of the NN model during equinoxes and solstices. Results show that NNs predict GPS TEC more accurately than the IRI at South African GPS locations, but that more good quality GPS data is required before a truly representative empirical GPS TEC model can be released.
- Full Text:
- Date Issued: 2008
Technology in conservation: towards a system for in-field drone detection of invasive vegetation
- Authors: James, Katherine Margaret Frances
- Date: 2020
- Subjects: Drone aircraft in remote sensing , Neural networks (Computer science) , Drone aircraft in remote sensing -- Case studies , Machine learning , Computer vision , Environmental monitoring -- Remote sensing , Invasive plants -- Monitoring
- Language: English
- Type: text , Thesis , Masters , MSc
- Identifier: http://hdl.handle.net/10962/143408 , vital:38244
- Description: Remote sensing can assist in monitoring the spread of invasive vegetation. The adoption of camera-carrying unmanned aerial vehicles, commonly referred to as drones, as remote sensing tools has yielded images of higher spatial resolution than traditional techniques. Drones also have the potential to interact with the environment through the delivery of bio-control or herbicide, as seen with their adoption in precision agriculture. Unlike in agricultural applications, however, invasive plants do not have a predictable position relative to each other within the environment. To facilitate the adoption of drones as an environmental monitoring and management tool, drones need to be able to intelligently distinguish between invasive and non-invasive vegetation on the fly. In this thesis, we present the augmentation of a commercially available drone with a deep machine learning model to investigate the viability of differentiating between an invasive shrub and other vegetation. As a case study, this was applied to the shrub genus Hakea, originating in Australia and invasive in several countries including South Africa. However, for this research, the methodology is important, rather than the chosen target plant. A dataset was collected using the available drone and manually annotated to facilitate the supervised training of the model. Two approaches were explored, namely, classification and semantic segmentation. For each of these, several models were trained and evaluated to find the optimal one. The chosen model was then interfaced with the drone via an Android application on a mobile device and its performance was preliminarily evaluated in the field. Based on these findings, refinements were made and thereafter a thorough field evaluation was performed to determine the best conditions for model operation. Results from the classification task show that deep learning models are capable of distinguishing between target and other shrubs in ideal candidate windows. However, classification in this manner is restricted by the proposal of such candidate windows. End-to-end image segmentation using deep learning overcomes this problem, classifying the image in a pixel-wise manner. Furthermore, the use of appropriate loss functions was found to improve model performance. Field tests show that illumination and shadow pose challenges to the model, but that good recall can be achieved when the conditions are ideal. False positive detection remains an issue that could be improved. This approach shows the potential for drones as an environmental monitoring and management tool when coupled with deep machine learning techniques and outlines potential problems that may be encountered.
- Full Text:
- Date Issued: 2020
Predictability of Geomagnetically Induced Currents using neural networks
- Authors: Lotz, Stefan
- Date: 2009
- Subjects: Advanced Composition Explorer (Artificial satellite) , Geomagnetism , Electromagnetic induction , Neural networks (Computer science) , Artificial intelligence
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5483 , http://hdl.handle.net/10962/d1005269 , Advanced Composition Explorer (Artificial satellite) , Geomagnetism , Electromagnetic induction , Neural networks (Computer science) , Artificial intelligence
- Description: It is a well documented fact that Geomagnetically Induced Currents (GIC’s) poses a significant threat to ground-based electric conductor networks like oil pipelines, railways and powerline networks. A study is undertaken to determine the feasibility of using artificial neural network models to predict GIC occurrence in the Southern African power grid. The magnitude of an induced current at a specific location on the Earth’s surface is directly related to the temporal derivative of the geomagnetic field (specifically its horizontal components) at that point. Hence, the focus of the problem is on the prediction of the temporal variations in the horizontal geomagnetic field (@Bx/@t and @By/@t). Artificial neural networks are used to predict @Bx/@t and @By/@t measured at Hermanus, South Africa (34.27◦ S, 19.12◦ E) with a 30 minute prediction lead time. As input parameters to the neural networks, insitu solar wind measurements made by the Advanced Composition Explorer (ACE) satellite are used. The results presented here compare well with similar models developed at high-latitude locations (e.g. Sweden, Finland, Canada) where extensive GIC research has been undertaken. It is concluded that it would indeed be feasible to use a neural network model to predict GIC occurrence in the Southern African power grid, provided that GIC measurements, powerline configuration and network parameters are made available.
- Full Text:
- Date Issued: 2009
Updating the ionospheric propagation factor, M(3000)F2, global model using the neural network technique and relevant geophysical input parameters
- Authors: Oronsaye, Samuel Iyen Jeffrey
- Date: 2013
- Subjects: Neural networks (Computer science) , Ionospheric radio wave propagation , Ionosphere , Geophysics , Ionosondes
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5434 , http://hdl.handle.net/10962/d1001609 , Neural networks (Computer science) , Ionospheric radio wave propagation , Ionosphere , Geophysics , Ionosondes
- Description: This thesis presents an update to the ionospheric propagation factor, M(3000)F2, global empirical model developed by Oyeyemi et al. (2007) (NNO). An additional aim of this research was to produce the updated model in a form that could be used within the International Reference Ionosphere (IRI) global model without adding to the complexity of the IRI. M(3000)F2 is the highest frequency at which a radio signal can be received over a distance of 3000 km after reflection in the ionosphere. The study employed the artificial neural network (ANN) technique using relevant geophysical input parameters which are known to influence the M(3000)F2 parameter. Ionosonde data from 135 ionospheric stations globally, including a number of equatorial stations, were available for this work. M(3000)F2 hourly values from 1976 to 2008, spanning all periods of low and high solar activity were used for model development and verification. A preliminary investigation was first carried out using a relatively small dataset to determine the appropriate input parameters for global M(3000)F2 parameter modelling. Inputs representing diurnal variation, seasonal variation, solar variation, modified dip latitude, longitude and latitude were found to be the optimum parameters for modelling the diurnal and seasonal variations of the M(3000)F2 parameter both on a temporal and spatial basis. The outcome of the preliminary study was applied to the overall dataset to develop a comprehensive ANN M(3000)F2 model which displays a remarkable improvement over the NNO model as well as the IRI version. The model shows 7.11% and 3.85% improvement over the NNO model as well as 13.04% and 10.05% over the IRI M(3000)F2 model, around high and low solar activity periods respectively. A comparison of the diurnal structure of the ANN and the IRI predicted values reveal that the ANN model is more effective in representing the diurnal structure of the M(3000)F2 values than the IRI M(3000)F2 model. The capability of the ANN model in reproducing the seasonal variation pattern of the M(3000)F2 values at 00h00UT, 06h00UT, 12h00UT, and l8h00UT more appropriately than the IRI version is illustrated in this work. A significant result obtained in this study is the ability of the ANN model in improving the post-sunset predicted values of the M(3000)F2 parameter which is known to be problematic to the IRI M(3000)F2 model in the low-latitude and the equatorial regions. The final M(3000)F2 model provides for an improved equatorial prediction and a simplified input space that allows for easy incorporation into the IRI model.
- Full Text:
- Date Issued: 2013
Development of a neural network based model for predicting the occurrence of spread F within the Brazilian sector
- Authors: Paradza, Masimba Wellington
- Date: 2009
- Subjects: Neural networks (Computer science) , Ionosphere , F region
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5460 , http://hdl.handle.net/10962/d1005245 , Neural networks (Computer science) , Ionosphere , F region
- Description: Spread F is a phenomenon of the ionosphere in which the pulses returned from the ionosphere are of a much greater duration than the transmitted ones. The occurrence of spread F can be predicted using the technique of Neural Networks (NNs). This thesis presents the development and evaluation of NN based models (two single station models and a regional model) for predicting the occurrence of spread F over selected stations within the Brazilian sector. The input space for the NNs included the day number (seasonal variation), hour (diurnal variation), sunspot number (measure of the solar activity), magnetic index (measure of the magnetic activity) and magnetic position (latitude, magnetic declination and inclination). Twelve years of spread F data measured during 1978 to 1989 inclusively at the equatorial site Fortaleza and low latitude site Cachoeira Paulista are used in the development of an input space and NN architecture for the NN models. Spread F data that is believed to be related to plasma bubble developments (range spread F) were used in the development of the models while those associated with narrow spectrum irregularities that occur near the F layer (frequency spread F) were excluded. The results of the models show the dependency of the probability of spread F as a function of local time, season and latitude. The models also illustrate some characteristics of spread F such as the onset and peak occurrence of spread F as a function of distance from the equator. Results from these models are presented in this thesis and compared to measured data and to modelled data obtained with an empirical model developed for the same purpose.
- Full Text:
- Date Issued: 2009
Artificial neural networks as simulators for behavioural evolution in evolutionary robotics
- Authors: Pretorius, Christiaan Johannes
- Date: 2010
- Subjects: Neural networks (Computer science) , Robotics
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:10462 , http://hdl.handle.net/10948/1476 , Neural networks (Computer science) , Robotics
- Description: Robotic simulators for use in Evolutionary Robotics (ER) have certain challenges associated with the complexity of their construction and the accuracy of predictions made by these simulators. Such robotic simulators are often based on physics models, which have been shown to produce accurate results. However, the construction of physics-based simulators can be complex and time-consuming. Alternative simulation schemes construct robotic simulators from empirically-collected data. Such empirical simulators, however, also have associated challenges, such as that some of these simulators do not generalize well on the data from which they are constructed, as these models employ simple interpolation on said data. As a result of the identified challenges in existing robotic simulators for use in ER, this project investigates the potential use of Artificial Neural Networks, henceforth simply referred to as Neural Networks (NNs), as alternative robotic simulators. In contrast to physics models, NN-based simulators can be constructed without needing an explicit mathematical model of the system being modeled, which can simplify simulator development. Furthermore, the generalization capabilities of NNs suggest that NNs could generalize well on data from which these simulators are constructed. These generalization abilities of NNs, along with NNs’ noise tolerance, suggest that NNs could be well-suited to application in robotics simulation. Investigating whether NNs can be effectively used as robotic simulators in ER is thus the endeavour of this work. Since not much research has been done in employing NNs as robotic simulators, many aspects of the experimental framework on which this dissertation reports needed to be carefully decided upon. Two robot morphologies were selected on which the NN simulators created in this work were based, namely a differentially steered robot and an inverted pendulum robot. Motion tracking and robotic sensor logging were used to acquire data from which the NN simulators were constructed. Furthermore, custom code was written for almost all aspects of the study, namely data acquisition for NN training, the actual NN training process, the evolution of robotic controllers using the created NN simulators, as well as the onboard robotic implementations of evolved controllers. Experimental tests performed in order to determine ideal topologies for each of the NN simulators developed in this study indicated that different NN topologies can lead to large differences in training accuracy. After performing these tests, the training accuracy of the created simulators was analyzed. This analysis showed that the NN simulators generally trained well and could generalize well on data not presented during simulator construction. In order to validate the feasibility of the created NN simulators in the ER process, these simulators were subsequently used to evolve controllers in simulation, similar to controllers developed in related studies. Encouraging results were obtained, with the newly-evolved controllers allowing real-world experimental robots to exhibit obstacle avoidance and light-approaching behaviour with a reasonable degree of success. The created NN simulators furthermore allowed for the successful evolution of a complex inverted pendulum stabilization controller in simulation. It was thus clearly established that NN-based robotic simulators can be successfully employed as alternative simulation schemes in the ER process.
- Full Text:
- Date Issued: 2010
Universal approximation properties of feedforward artificial neural networks.
- Authors: Redpath, Stuart Frederick
- Date: 2011
- Subjects: Neural networks (Computer science) , Artificial intelligence -- Biological applications , Functional analysis , Weierstrass-Stone Theorem , Banach-Hahn theorem
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5430 , http://hdl.handle.net/10962/d1015869
- Description: In this thesis we summarise several results in the literature which show the approximation capabilities of multilayer feedforward artificial neural networks. We show that multilayer feedforward artificial neural networks are capable of approximating continuous and measurable functions from Rn to R to any degree of accuracy under certain conditions. In particular making use of the Stone-Weierstrass and Hahn-Banach theorems, we show that a multilayer feedforward artificial neural network can approximate any continuous function to any degree of accuracy, by using either an arbitrary squashing function or any continuous sigmoidal function for activation. Making use of the Stone-Weirstrass Theorem again, we extend these approximation capabilities of multilayer feedforward artificial neural networks to the space of measurable functions under any probability measure.
- Full Text:
- Date Issued: 2011
Protein secondary structure prediction using neural networks and support vector machines
- Authors: Tsilo, Lipontseng Cecilia
- Date: 2009
- Subjects: Neural networks (Computer science) , Support vector machines , Proteins -- Structure -- Mathematical models
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5569 , http://hdl.handle.net/10962/d1002809 , Neural networks (Computer science) , Support vector machines , Proteins -- Structure -- Mathematical models
- Description: Predicting the secondary structure of proteins is important in biochemistry because the 3D structure can be determined from the local folds that are found in secondary structures. Moreover, knowing the tertiary structure of proteins can assist in determining their functions. The objective of this thesis is to compare the performance of Neural Networks (NN) and Support Vector Machines (SVM) in predicting the secondary structure of 62 globular proteins from their primary sequence. For each NN and SVM, we created six binary classifiers to distinguish between the classes’ helices (H) strand (E), and coil (C). For NN we use Resilient Backpropagation training with and without early stopping. We use NN with either no hidden layer or with one hidden layer with 1,2,...,40 hidden neurons. For SVM we use a Gaussian kernel with parameter fixed at = 0.1 and varying cost parameters C in the range [0.1,5]. 10- fold cross-validation is used to obtain overall estimates for the probability of making a correct prediction. Our experiments indicate for NN and SVM that the different binary classifiers have varying accuracies: from 69% correct predictions for coils vs. non-coil up to 80% correct predictions for stand vs. non-strand. It is further demonstrated that NN with no hidden layer or not more than 2 hidden neurons in the hidden layer are sufficient for better predictions. For SVM we show that the estimated accuracies do not depend on the value of the cost parameter. As a major result, we will demonstrate that the accuracy estimates of NN and SVM binary classifiers cannot distinguish. This contradicts a modern belief in bioinformatics that SVM outperforms other predictors.
- Full Text:
- Date Issued: 2009
Forecasting solar cycle 24 using neural networks
- Authors: Uwamahoro, Jean
- Date: 2009
- Subjects: Solar cycle , Neural networks (Computer science) , Ionosphere , Ionospheric electron density , Ionospheric forecasting , Solar thermal energy
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5468 , http://hdl.handle.net/10962/d1005253 , Solar cycle , Neural networks (Computer science) , Ionosphere , Ionospheric electron density , Ionospheric forecasting , Solar thermal energy
- Description: The ability to predict the future behavior of solar activity has become of extreme importance due to its effect on the near-Earth environment. Predictions of both the amplitude and timing of the next solar cycle will assist in estimating the various consequences of Space Weather. Several prediction techniques have been applied and have achieved varying degrees of success in the domain of solar activity prediction. These techniques include, for example, neural networks and geomagnetic precursor methods. In this thesis, various neural network based models were developed and the model considered to be optimum was used to estimate the shape and timing of solar cycle 24. Given the recent success of the geomagnetic precusrsor methods, geomagnetic activity as measured by the aa index is considered among the main inputs to the neural network model. The neural network model developed is also provided with the time input parameters defining the year and the month of a particular solar cycle, in order to characterise the temporal behaviour of sunspot number as observed during the last 10 solar cycles. The structure of input-output patterns to the neural network is constructed in such a way that the network learns the relationship between the aa index values of a particular cycle, and the sunspot number values of the following cycle. Assuming January 2008 as the minimum preceding solar cycle 24, the shape and amplitude of solar cycle 24 is estimated in terms of monthly mean and smoothed monthly sunspot number. This new prediction model estimates an average solar cycle 24, with the maximum occurring around June 2012 [± 11 months], with a smoothed monthly maximum sunspot number of 121 ± 9.
- Full Text:
- Date Issued: 2009