Improving licence plate detection using generative adversarial networks
- Authors: Boby, Alden , Brown, Dane L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/464145 , vital:76480 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-031-04881-4_47"
- Description: The information on a licence plate is used for traffic law enforcement, access control, surveillance and parking lot management. Existing li-cence plate recognition systems work with clear images taken under controlled conditions. In real-world licence plate recognition scenarios, images are not as straightforward as the ‘toy’ datasets used to bench-mark existing systems. Real-world data is often noisy as it may contain occlusion and poor lighting, obscuring the information on a licence plate. Cleaning input data before using it for licence plate recognition is a complex problem, and existing literature addressing the issue is still limited. This paper uses two deep learning techniques to improve li-cence plate visibility towards more accurate licence plate recognition. A one-stage object detector popularly known as YOLO is implemented for locating licence plates under challenging situations. Super-resolution generative adversarial networks are considered for image upscaling and reconstruction to improve the clarity of low-quality input. The main focus involves training these systems on datasets that include difficult to detect licence plates, enabling better performance in unfavourable conditions and environments.
- Full Text:
- Date Issued: 2022
- Authors: Boby, Alden , Brown, Dane L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/464145 , vital:76480 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-031-04881-4_47"
- Description: The information on a licence plate is used for traffic law enforcement, access control, surveillance and parking lot management. Existing li-cence plate recognition systems work with clear images taken under controlled conditions. In real-world licence plate recognition scenarios, images are not as straightforward as the ‘toy’ datasets used to bench-mark existing systems. Real-world data is often noisy as it may contain occlusion and poor lighting, obscuring the information on a licence plate. Cleaning input data before using it for licence plate recognition is a complex problem, and existing literature addressing the issue is still limited. This paper uses two deep learning techniques to improve li-cence plate visibility towards more accurate licence plate recognition. A one-stage object detector popularly known as YOLO is implemented for locating licence plates under challenging situations. Super-resolution generative adversarial networks are considered for image upscaling and reconstruction to improve the clarity of low-quality input. The main focus involves training these systems on datasets that include difficult to detect licence plates, enabling better performance in unfavourable conditions and environments.
- Full Text:
- Date Issued: 2022
Improving signer-independence using pose estimation and transfer learning for sign language recognition
- Marais, Marc, Brown, Dane L, Connan, James, Boby, Alden
- Authors: Marais, Marc , Brown, Dane L , Connan, James , Boby, Alden
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/463406 , vital:76406 , xlink:href="https://doi.org/10.1007/978-3-031-35644-5"
- Description: Automated Sign Language Recognition (SLR) aims to bridge the com-munication gap between the hearing and the hearing disabled. Com-puter vision and deep learning lie at the forefront in working toward these systems. Most SLR research focuses on signer-dependent SLR and fails to account for variations in varying signers who gesticulate naturally. This paper investigates signer-independent SLR on the LSA64 dataset, focusing on different feature extraction approaches. Two approaches are proposed an InceptionV3-GRU architecture, which uses raw images as input, and a pose estimation LSTM architecture. MediaPipe Holistic is implemented to extract pose estimation landmark coordinates. A final third model applies augmentation and transfer learning using the pose estimation LSTM model. The research found that the pose estimation LSTM approach achieved the best perfor-mance with an accuracy of 80.22%. MediaPipe Holistic struggled with the augmentations introduced in the final experiment. Thus, looking into introducing more subtle augmentations may improve the model. Over-all, the system shows significant promise toward addressing the real-world signer-independence issue in SLR.
- Full Text:
- Date Issued: 2022
- Authors: Marais, Marc , Brown, Dane L , Connan, James , Boby, Alden
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/463406 , vital:76406 , xlink:href="https://doi.org/10.1007/978-3-031-35644-5"
- Description: Automated Sign Language Recognition (SLR) aims to bridge the com-munication gap between the hearing and the hearing disabled. Com-puter vision and deep learning lie at the forefront in working toward these systems. Most SLR research focuses on signer-dependent SLR and fails to account for variations in varying signers who gesticulate naturally. This paper investigates signer-independent SLR on the LSA64 dataset, focusing on different feature extraction approaches. Two approaches are proposed an InceptionV3-GRU architecture, which uses raw images as input, and a pose estimation LSTM architecture. MediaPipe Holistic is implemented to extract pose estimation landmark coordinates. A final third model applies augmentation and transfer learning using the pose estimation LSTM model. The research found that the pose estimation LSTM approach achieved the best perfor-mance with an accuracy of 80.22%. MediaPipe Holistic struggled with the augmentations introduced in the final experiment. Thus, looking into introducing more subtle augmentations may improve the model. Over-all, the system shows significant promise toward addressing the real-world signer-independence issue in SLR.
- Full Text:
- Date Issued: 2022
Investigating signer-independent sign language recognition on the lsa64 dataset
- Marais, Marc, Brown, Dane L, Connan, James, Boby, Alden, Kuhlane, Luxolo L
- Authors: Marais, Marc , Brown, Dane L , Connan, James , Boby, Alden , Kuhlane, Luxolo L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465179 , vital:76580 , xlink:href="https://www.researchgate.net/profile/Marc-Marais/publication/363174384_Investigating_Signer-Independ-ent_Sign_Language_Recognition_on_the_LSA64_Dataset/links/63108c7d5eed5e4bd138680f/Investigating-Signer-Independent-Sign-Language-Recognition-on-the-LSA64-Dataset.pdf"
- Description: Conversing with hearing disabled people is a significant challenge; however, computer vision advancements have significantly improved this through automated sign language recognition. One of the common issues in sign language recognition is signer-dependence, where variations arise from varying signers, who gesticulate naturally. Utilising the LSA64 dataset, a small scale Argentinian isolated sign language recognition, we investigate signer-independent sign language recognition. An InceptionV3-GRU architecture is employed to extract and classify spatial and temporal information for automated sign language recognition. The signer-dependent approach yielded an accuracy of 97.03%, whereas the signer-independent approach achieved an accuracy of 74.22%. The signer-independent system shows promise towards addressing the real-world and common issue of signer-dependence in sign language recognition.
- Full Text:
- Date Issued: 2022
- Authors: Marais, Marc , Brown, Dane L , Connan, James , Boby, Alden , Kuhlane, Luxolo L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465179 , vital:76580 , xlink:href="https://www.researchgate.net/profile/Marc-Marais/publication/363174384_Investigating_Signer-Independ-ent_Sign_Language_Recognition_on_the_LSA64_Dataset/links/63108c7d5eed5e4bd138680f/Investigating-Signer-Independent-Sign-Language-Recognition-on-the-LSA64-Dataset.pdf"
- Description: Conversing with hearing disabled people is a significant challenge; however, computer vision advancements have significantly improved this through automated sign language recognition. One of the common issues in sign language recognition is signer-dependence, where variations arise from varying signers, who gesticulate naturally. Utilising the LSA64 dataset, a small scale Argentinian isolated sign language recognition, we investigate signer-independent sign language recognition. An InceptionV3-GRU architecture is employed to extract and classify spatial and temporal information for automated sign language recognition. The signer-dependent approach yielded an accuracy of 97.03%, whereas the signer-independent approach achieved an accuracy of 74.22%. The signer-independent system shows promise towards addressing the real-world and common issue of signer-dependence in sign language recognition.
- Full Text:
- Date Issued: 2022
Investigating the Effects of Image Correction Through Affine Transformations on Licence Plate Recognition
- Boby, Alden, Brown, Dane L, Connan, James, Marais, Marc
- Authors: Boby, Alden , Brown, Dane L , Connan, James , Marais, Marc
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465190 , vital:76581 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9856380"
- Description: Licence plate recognition has many real-world applications, which fall under security and surveillance. Deep learning for licence plate recognition has been adopted to improve existing image-based processing techniques in recent years. Object detectors are a popular choice for approaching this task. All object detectors are some form of a convolutional neural network. The You Only Look Once framework and Region-Based Convolutional Neural Networks are popular models within this field. A novel architecture called the Warped Planar Object Detector is a recent development by Zou et al. that takes inspiration from YOLO and Spatial Network Transformers. This paper aims to compare the performance of the Warped Planar Object Detector and YOLO on licence plate recognition by training both models with the same data and then directing their output to an Enhanced Super-Resolution Generative Adversarial Network to upscale the output image, then lastly using an Optical Character Recognition engine to classify characters detected from the images.
- Full Text:
- Date Issued: 2022
- Authors: Boby, Alden , Brown, Dane L , Connan, James , Marais, Marc
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465190 , vital:76581 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9856380"
- Description: Licence plate recognition has many real-world applications, which fall under security and surveillance. Deep learning for licence plate recognition has been adopted to improve existing image-based processing techniques in recent years. Object detectors are a popular choice for approaching this task. All object detectors are some form of a convolutional neural network. The You Only Look Once framework and Region-Based Convolutional Neural Networks are popular models within this field. A novel architecture called the Warped Planar Object Detector is a recent development by Zou et al. that takes inspiration from YOLO and Spatial Network Transformers. This paper aims to compare the performance of the Warped Planar Object Detector and YOLO on licence plate recognition by training both models with the same data and then directing their output to an Enhanced Super-Resolution Generative Adversarial Network to upscale the output image, then lastly using an Optical Character Recognition engine to classify characters detected from the images.
- Full Text:
- Date Issued: 2022
Iterative Refinement Versus Generative Adversarial Networks for Super-Resolution Towards Licence Plate Detection
- Boby, Alden, Brown, Dane L, Connan, James
- Authors: Boby, Alden , Brown, Dane L , Connan, James
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/463417 , vital:76407 , xlink:href="https://link.springer.com/chapter/10.1007/978-981-99-1624-5_26"
- Description: Licence plate detection in unconstrained scenarios can be difficult because of the medium used to capture the data. Such data is not captured at very high resolution for practical reasons. Super-resolution can be used to improve the resolution of an image with fidelity beyond that of non-machine learning-based image upscaling algorithms such as bilinear or bicubic upscaling. Technological advances have introduced more than one way to perform super-resolution, with the best results coming from generative adversarial networks and iterative refinement with diffusion-based models. This paper puts the two best-performing super-resolution models against each other to see which is best for licence plate super-resolution. Quantitative results favour the generative adversarial network, while qualitative results lean towards the iterative refinement model.
- Full Text:
- Date Issued: 2022
- Authors: Boby, Alden , Brown, Dane L , Connan, James
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/463417 , vital:76407 , xlink:href="https://link.springer.com/chapter/10.1007/978-981-99-1624-5_26"
- Description: Licence plate detection in unconstrained scenarios can be difficult because of the medium used to capture the data. Such data is not captured at very high resolution for practical reasons. Super-resolution can be used to improve the resolution of an image with fidelity beyond that of non-machine learning-based image upscaling algorithms such as bilinear or bicubic upscaling. Technological advances have introduced more than one way to perform super-resolution, with the best results coming from generative adversarial networks and iterative refinement with diffusion-based models. This paper puts the two best-performing super-resolution models against each other to see which is best for licence plate super-resolution. Quantitative results favour the generative adversarial network, while qualitative results lean towards the iterative refinement model.
- Full Text:
- Date Issued: 2022
Plant disease detection using deep learning on natural environment images
- De Silva, Malitha, Brown, Dane L
- Authors: De Silva, Malitha , Brown, Dane L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465212 , vital:76583 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9855925"
- Description: Improving agriculture is one of the major concerns today, as it helps reduce global hunger. In past years, many technological advancements have been introduced to enhance harvest quality and quantity by controlling and preventing weeds, pests, and diseases. Several studies have focused on identifying diseases in plants, as it helps to make decisions on spraying fungicides and fertilizers. State-of-the-art systems typically combine image processing and deep learning methods to identify conditions with visible symptoms. However, they use already available data sets or images taken in controlled environments. This study was conducted on two data sets of ten plants collected in a natural environment. The first dataset contained RGB Visible images, while the second contained Near-Infrared (NIR) images of healthy and diseased leaves. The visible image dataset showed higher training and validation accuracies than the NIR image dataset with ResNet, Inception, VGG and MobileNet architectures. For the visible image and NIR dataset, ResNet-50V2 outperformed other models with validation accuracies of 98.35% and 94.01%, respectively.
- Full Text:
- Date Issued: 2022
- Authors: De Silva, Malitha , Brown, Dane L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465212 , vital:76583 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9855925"
- Description: Improving agriculture is one of the major concerns today, as it helps reduce global hunger. In past years, many technological advancements have been introduced to enhance harvest quality and quantity by controlling and preventing weeds, pests, and diseases. Several studies have focused on identifying diseases in plants, as it helps to make decisions on spraying fungicides and fertilizers. State-of-the-art systems typically combine image processing and deep learning methods to identify conditions with visible symptoms. However, they use already available data sets or images taken in controlled environments. This study was conducted on two data sets of ten plants collected in a natural environment. The first dataset contained RGB Visible images, while the second contained Near-Infrared (NIR) images of healthy and diseased leaves. The visible image dataset showed higher training and validation accuracies than the NIR image dataset with ResNet, Inception, VGG and MobileNet architectures. For the visible image and NIR dataset, ResNet-50V2 outperformed other models with validation accuracies of 98.35% and 94.01%, respectively.
- Full Text:
- Date Issued: 2022
Plant disease detection using multispectral imaging
- De Silva, Malitha, Brown, Dane L
- Authors: De Silva, Malitha , Brown, Dane L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/463439 , vital:76409 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-031-35641-4_24"
- Description: People worldwide are undergoing many challenges, including food scarcity. Many pieces of research are now focused on improving agriculture to increase the harvest and reduce the cost. Identifying plant diseases and pests in the early stages helps to enhance the yield and reduce costs. However, most plant disease identification research with computer vision has been done with images taken in controlled environments on publically available data sets. Near-Infrared (NIR) imaging is a favourable approach for identifying plant diseases. Therefore, this study collected NIR images of healthy and diseased leaves in the natural environment. The dataset is tested with eight Convolutional Neural Network (CNN) models with different train-test splits ranging from 10:90 to 90:10. The evaluated models attained their highest training and test accuracies from the 70:30 split onwards. Xception outperformed all the other models in all train-test splits and achieved 100% accuracy, precision and recall in the 80:20 train-test split.
- Full Text:
- Date Issued: 2022
- Authors: De Silva, Malitha , Brown, Dane L
- Date: 2022
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/463439 , vital:76409 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-031-35641-4_24"
- Description: People worldwide are undergoing many challenges, including food scarcity. Many pieces of research are now focused on improving agriculture to increase the harvest and reduce the cost. Identifying plant diseases and pests in the early stages helps to enhance the yield and reduce costs. However, most plant disease identification research with computer vision has been done with images taken in controlled environments on publically available data sets. Near-Infrared (NIR) imaging is a favourable approach for identifying plant diseases. Therefore, this study collected NIR images of healthy and diseased leaves in the natural environment. The dataset is tested with eight Convolutional Neural Network (CNN) models with different train-test splits ranging from 10:90 to 90:10. The evaluated models attained their highest training and test accuracies from the 70:30 split onwards. Xception outperformed all the other models in all train-test splits and achieved 100% accuracy, precision and recall in the 80:20 train-test split.
- Full Text:
- Date Issued: 2022
A Robust Portable Environment for First-Year Computer Science Students
- Brown, Dane L, Connan, James
- Authors: Brown, Dane L , Connan, James
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465113 , vital:76574 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-030-92858-2_6"
- Description: Computer science education in both South African universities and worldwide often aim at making students confident at problem solving by introducing various programming exercises. Standardising a computer environment where students can apply their computational thinking knowledge on a more even playing field – without worrying about software issues – can be beneficial for problem solving in classroom of diverse students. Research shows that having consistent access to this exposes students to core concepts of Computer Science. However, with the diverse student base in South Africa, not everyone has access to a personal computer or expensive software. This paper describes a new approach at first-year level that uses the power of a modified Linux distro on a flash drive to enable access to the same, fully-fledged, free and open-source environment, including the convenience of portability. This is used as a means to even the playing field in a diverse country like South Africa and address the lack of consistent access to a problem solving environment. Feedback from students and staff at the Institution are effectively heeded and attempted to be measured.
- Full Text:
- Date Issued: 2021
- Authors: Brown, Dane L , Connan, James
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465113 , vital:76574 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-030-92858-2_6"
- Description: Computer science education in both South African universities and worldwide often aim at making students confident at problem solving by introducing various programming exercises. Standardising a computer environment where students can apply their computational thinking knowledge on a more even playing field – without worrying about software issues – can be beneficial for problem solving in classroom of diverse students. Research shows that having consistent access to this exposes students to core concepts of Computer Science. However, with the diverse student base in South Africa, not everyone has access to a personal computer or expensive software. This paper describes a new approach at first-year level that uses the power of a modified Linux distro on a flash drive to enable access to the same, fully-fledged, free and open-source environment, including the convenience of portability. This is used as a means to even the playing field in a diverse country like South Africa and address the lack of consistent access to a problem solving environment. Feedback from students and staff at the Institution are effectively heeded and attempted to be measured.
- Full Text:
- Date Issued: 2021
Adaptive machine learning based network intrusion detection
- Chindove, Hatitye E, Brown, Dane L
- Authors: Chindove, Hatitye E , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/464052 , vital:76471 , xlink:href="https://doi.org/10.1145/3487923.3487938"
- Description: Network intrusion detection system (NIDS) adoption is essential for mitigating computer network attacks in various scenarios. However, the increasing complexity of computer networks and attacks make it challenging to classify network traffic. Machine learning (ML) techniques in a NIDS can be affected by different scenarios, and thus the recency, size and applicability of datasets are vital factors to consider when selecting and tuning a machine learning classifier. The proposed approach evaluates relatively new datasets constructed such that they depict real-world scenarios. It includes analyses of dataset balancing and sampling, feature engineering and systematic ML-based NIDS model tuning focused on the adaptive improvement of intrusion detection. A comparison between machine learning classifiers forms part of the evaluation process. Results on the proposed approach model effectiveness for NIDS are discussed. Recurrent neural networks and random forests models consistently achieved high f1-score results with macro f1-scores of 0.73 and 0.87 for the CICIDS 2017 dataset; and 0.73 and 0.72 against the CICIDS 2018 dataset, respectively.
- Full Text:
- Date Issued: 2021
- Authors: Chindove, Hatitye E , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/464052 , vital:76471 , xlink:href="https://doi.org/10.1145/3487923.3487938"
- Description: Network intrusion detection system (NIDS) adoption is essential for mitigating computer network attacks in various scenarios. However, the increasing complexity of computer networks and attacks make it challenging to classify network traffic. Machine learning (ML) techniques in a NIDS can be affected by different scenarios, and thus the recency, size and applicability of datasets are vital factors to consider when selecting and tuning a machine learning classifier. The proposed approach evaluates relatively new datasets constructed such that they depict real-world scenarios. It includes analyses of dataset balancing and sampling, feature engineering and systematic ML-based NIDS model tuning focused on the adaptive improvement of intrusion detection. A comparison between machine learning classifiers forms part of the evaluation process. Results on the proposed approach model effectiveness for NIDS are discussed. Recurrent neural networks and random forests models consistently achieved high f1-score results with macro f1-scores of 0.73 and 0.87 for the CICIDS 2017 dataset; and 0.73 and 0.72 against the CICIDS 2018 dataset, respectively.
- Full Text:
- Date Issued: 2021
Adaptive network intrusion detection using optimised machine learning models
- Chindove, Hatitye E, Brown, Dane L
- Authors: Chindove, Hatitye E , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465634 , vital:76627 , xlink:href="https://www.researchgate.net/publication/358046953_Adaptive_Network_Intrusion_Detection_using_Optimised_Machine_Learning_Models"
- Description: Network intrusion detection system (NIDS) adoption is essential for mitigating computer network attacks in various scenarios. However, the increasing complexity of computer networks and attacks make it challenging to classify network traffic. Machine learning (ML) techniques in a NIDS can be affected by different scenarios, and thus the recency, size and applicability of datasets are vital factors to consider when selecting and tuning a machine learning classifier. The proposed approach evaluates relatively new datasets constructed such that they depict real-world scenarios. It includes empirical analyses of practical, systematic ML-based NIDS with significant network traffic for improved intrusion detection. A comparison between machine learning classifiers, including deep learning, form part of the evaluation process. Results on how the proposed approach increased model effectiveness for NIDS in a more practical setting are discussed. Recurrent neural networks and random forests models consistently achieved the best results.
- Full Text:
- Date Issued: 2021
- Authors: Chindove, Hatitye E , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465634 , vital:76627 , xlink:href="https://www.researchgate.net/publication/358046953_Adaptive_Network_Intrusion_Detection_using_Optimised_Machine_Learning_Models"
- Description: Network intrusion detection system (NIDS) adoption is essential for mitigating computer network attacks in various scenarios. However, the increasing complexity of computer networks and attacks make it challenging to classify network traffic. Machine learning (ML) techniques in a NIDS can be affected by different scenarios, and thus the recency, size and applicability of datasets are vital factors to consider when selecting and tuning a machine learning classifier. The proposed approach evaluates relatively new datasets constructed such that they depict real-world scenarios. It includes empirical analyses of practical, systematic ML-based NIDS with significant network traffic for improved intrusion detection. A comparison between machine learning classifiers, including deep learning, form part of the evaluation process. Results on how the proposed approach increased model effectiveness for NIDS in a more practical setting are discussed. Recurrent neural networks and random forests models consistently achieved the best results.
- Full Text:
- Date Issued: 2021
An Evaluation of Machine Learning Methods for Classifying Bot Traffic in Software Defined Networks
- Van Staden, Joshua, Brown, Dane L
- Authors: Van Staden, Joshua , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465645 , vital:76628 , xlink:href="https://link.springer.com/chapter/10.1007/978-981-19-7874-6_72"
- Description: Internet security is an ever-expanding field. Cyber-attacks occur very frequently, and so detecting them is an important aspect of preserving services. Machine learning offers a helpful tool with which to detect cyber attacks. However, it is impossible to deploy a machine-learning algorithm to detect attacks in a non-centralized network. Software Defined Networks (SDNs) offer a centralized view of a network, allowing machine learning algorithms to detect malicious activity within a network. The InSDN dataset is a recently-released dataset that contains a set of sniffed packets within a virtual SDN. These sniffed packets correspond to various attacks, including DDoS attacks, Probing and Password-Guessing, among others. This study aims to evaluate various machine learning models against this new dataset. Specifically, we aim to evaluate their classification ability and runtimes when trained on fewer features. The machine learning models tested include a Neural Network, Support Vector Machine, Random Forest, Multilayer Perceptron, Logistic Regression, and K-Nearest Neighbours. Cluster-based algorithms such as the K-Nearest Neighbour and Random Forest proved to be the best performers. Linear-based algorithms such as the Multilayer Perceptron performed the worst. This suggests a good level of clustering in the top few features with little space for linear separability. The reduction of features significantly reduced training time, particularly in the better-performing models.
- Full Text:
- Date Issued: 2021
- Authors: Van Staden, Joshua , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465645 , vital:76628 , xlink:href="https://link.springer.com/chapter/10.1007/978-981-19-7874-6_72"
- Description: Internet security is an ever-expanding field. Cyber-attacks occur very frequently, and so detecting them is an important aspect of preserving services. Machine learning offers a helpful tool with which to detect cyber attacks. However, it is impossible to deploy a machine-learning algorithm to detect attacks in a non-centralized network. Software Defined Networks (SDNs) offer a centralized view of a network, allowing machine learning algorithms to detect malicious activity within a network. The InSDN dataset is a recently-released dataset that contains a set of sniffed packets within a virtual SDN. These sniffed packets correspond to various attacks, including DDoS attacks, Probing and Password-Guessing, among others. This study aims to evaluate various machine learning models against this new dataset. Specifically, we aim to evaluate their classification ability and runtimes when trained on fewer features. The machine learning models tested include a Neural Network, Support Vector Machine, Random Forest, Multilayer Perceptron, Logistic Regression, and K-Nearest Neighbours. Cluster-based algorithms such as the K-Nearest Neighbour and Random Forest proved to be the best performers. Linear-based algorithms such as the Multilayer Perceptron performed the worst. This suggests a good level of clustering in the top few features with little space for linear separability. The reduction of features significantly reduced training time, particularly in the better-performing models.
- Full Text:
- Date Issued: 2021
An Evaluation of YOLO-Based Algorithms for Hand Detection in the Kitchen
- Van Staden, Joshua, Brown, Dane L
- Authors: Van Staden, Joshua , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465134 , vital:76576 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9519307"
- Description: Convolutional Neural Networks have offered an accurate method with which to run object detection on images. Specifically, the YOLO family of object detection algorithms have proven to be relatively fast and accurate. Since its inception, the different variants of this algorithm have been tested on different datasets. In this paper, we evaluate the performances of these algorithms on the recent Epic Kitchens-100 dataset. This dataset provides egocentric footage of people interacting with various objects in the kitchen. Most prominently shown in the footage is an egocentric view of the participants' hands. We aim to use the YOLOv3 algorithm to detect these hands within the footage provided in this dataset. In particular, we examine the YOLOv3 algorithm using two different backbones: MobileNet-lite and VGG16. We trained them on a mixture of samples from the Egohands and Epic Kitchens-100 datasets. In a separate experiment, average precision was measured on an unseen Epic Kitchens-100 subset. We found that the models are relatively simple and lead to lower scores on the Epic Kitchens 100 dataset. This is attributed to the high background noise on the Epic Kitchens 100 dataset. Nonetheless, the VGG16 architecture was found to have a higher Average Precision (AP) and is, therefore, more suited for retrospective analysis. None of the models was suitable for real-time analysis due to complex egocentric data.
- Full Text:
- Date Issued: 2021
- Authors: Van Staden, Joshua , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465134 , vital:76576 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9519307"
- Description: Convolutional Neural Networks have offered an accurate method with which to run object detection on images. Specifically, the YOLO family of object detection algorithms have proven to be relatively fast and accurate. Since its inception, the different variants of this algorithm have been tested on different datasets. In this paper, we evaluate the performances of these algorithms on the recent Epic Kitchens-100 dataset. This dataset provides egocentric footage of people interacting with various objects in the kitchen. Most prominently shown in the footage is an egocentric view of the participants' hands. We aim to use the YOLOv3 algorithm to detect these hands within the footage provided in this dataset. In particular, we examine the YOLOv3 algorithm using two different backbones: MobileNet-lite and VGG16. We trained them on a mixture of samples from the Egohands and Epic Kitchens-100 datasets. In a separate experiment, average precision was measured on an unseen Epic Kitchens-100 subset. We found that the models are relatively simple and lead to lower scores on the Epic Kitchens 100 dataset. This is attributed to the high background noise on the Epic Kitchens 100 dataset. Nonetheless, the VGG16 architecture was found to have a higher Average Precision (AP) and is, therefore, more suited for retrospective analysis. None of the models was suitable for real-time analysis due to complex egocentric data.
- Full Text:
- Date Issued: 2021
Early dehydration detection using infrared imaging
- Poole, Louise C, Brown, Dane L, Connan, James
- Authors: Poole, Louise C , Brown, Dane L , Connan, James
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465656 , vital:76629 , xlink:href="https://www.researchgate.net/profile/Louise-Poole-3/publication/357578445_Early_Dehydration_Detection_Using_Infrared_Imaging/links/61d5664eb8305f7c4b231d50/Early-Dehydration-Detection-Using-Infrared-Imaging.pdf"
- Description: Crop loss and failure have devastating impacts on a country’s economy and food security. Developing effective and inexpensive systems to minimize crop loss has become essential. Recently, multispectral imaging—in particular visible and infrared imaging—have become popular for analyzing plants and show potential for early identification of plant stress. We created a directly comparable visible and infrared image dataset for dehydration in spinach leaves. We created and compared various models trained on both datasets and concluded that the models trained on the infrared dataset outperformed all of those trained on the visible dataset. In particular, the models trained to identify early signs of dehydration yielded 45% difference in accuracy, with the infrared model obtaining 70% accuracy and the visible model obtaining 25% accuracy. Infrared imaging thus shows promising potential for application in early plant stress and disease identification.
- Full Text:
- Date Issued: 2021
- Authors: Poole, Louise C , Brown, Dane L , Connan, James
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465656 , vital:76629 , xlink:href="https://www.researchgate.net/profile/Louise-Poole-3/publication/357578445_Early_Dehydration_Detection_Using_Infrared_Imaging/links/61d5664eb8305f7c4b231d50/Early-Dehydration-Detection-Using-Infrared-Imaging.pdf"
- Description: Crop loss and failure have devastating impacts on a country’s economy and food security. Developing effective and inexpensive systems to minimize crop loss has become essential. Recently, multispectral imaging—in particular visible and infrared imaging—have become popular for analyzing plants and show potential for early identification of plant stress. We created a directly comparable visible and infrared image dataset for dehydration in spinach leaves. We created and compared various models trained on both datasets and concluded that the models trained on the infrared dataset outperformed all of those trained on the visible dataset. In particular, the models trained to identify early signs of dehydration yielded 45% difference in accuracy, with the infrared model obtaining 70% accuracy and the visible model obtaining 25% accuracy. Infrared imaging thus shows promising potential for application in early plant stress and disease identification.
- Full Text:
- Date Issued: 2021
Handwriting Recognition using Deep Learning with Effective Data Augmentation Techniques
- Brown, Dane L, Lidzhade, Ipfi
- Authors: Brown, Dane L , Lidzhade, Ipfi
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465700 , vital:76633 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9519359"
- Description: Machine learning techniques have been successfully used in deciphering handwritten text. Deep learning has made further improvements in this regard. However, they require substantial amounts of training data. This research aims to improve the effectiveness of classification accuracy in the presence of limited training data on handwriting recognition. The main focus thus involves enabling deep models to converge during training on smaller datasets using data augmentation. This will allow for broader use of these systems across more regions, greater accessibility, and future related systems to be less reliant on the amount of data available. Therefore, the proposed research includes an image processing and machine learning approach to handwriting recognition while generating more sample data in various ways. Applying random cropping as an augmentation technique resulted in higher accuracy than several other augmentation techniques examined in this paper. Some of these techniques performed worse than on unaugmented data.
- Full Text:
- Date Issued: 2021
- Authors: Brown, Dane L , Lidzhade, Ipfi
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465700 , vital:76633 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9519359"
- Description: Machine learning techniques have been successfully used in deciphering handwritten text. Deep learning has made further improvements in this regard. However, they require substantial amounts of training data. This research aims to improve the effectiveness of classification accuracy in the presence of limited training data on handwriting recognition. The main focus thus involves enabling deep models to converge during training on smaller datasets using data augmentation. This will allow for broader use of these systems across more regions, greater accessibility, and future related systems to be less reliant on the amount of data available. Therefore, the proposed research includes an image processing and machine learning approach to handwriting recognition while generating more sample data in various ways. Applying random cropping as an augmentation technique resulted in higher accuracy than several other augmentation techniques examined in this paper. Some of these techniques performed worse than on unaugmented data.
- Full Text:
- Date Issued: 2021
Investigating popular CNN architectures for plant disease detection
- Poole, Louise C, Brown, Dane L
- Authors: Poole, Louise C , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465168 , vital:76579 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9519341"
- Description: Food production and food security have become increasingly important due to climate change and rising population numbers. One method to prevent crop loss is to develop a system to allow for early, efficient and accurate identification of plant diseases. CNNs often outperform previously popular machine learning algorithms. There are many existing CNN architectures. We compared and analysed the popular state-of-the-art architectures, namely ResNet, GoogLeNet and VGG, when trained for plant disease classification. We found that ResNet performed the best on the balanced Mendeley Leaves and PlantVillage datasets, obtaining 91.95% and 95.80% accuracy respectively. However, the ResNet architecture was relatively computationally expensive and slow to train. GoogLeNet obtained accuracies very close to those of ResNet with 89.35% and 94.59% achieved on the Mendeley Leaves and PlantVillage datasets respectively and could be considered a less computationally expensive alternative.
- Full Text:
- Date Issued: 2021
- Authors: Poole, Louise C , Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465168 , vital:76579 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9519341"
- Description: Food production and food security have become increasingly important due to climate change and rising population numbers. One method to prevent crop loss is to develop a system to allow for early, efficient and accurate identification of plant diseases. CNNs often outperform previously popular machine learning algorithms. There are many existing CNN architectures. We compared and analysed the popular state-of-the-art architectures, namely ResNet, GoogLeNet and VGG, when trained for plant disease classification. We found that ResNet performed the best on the balanced Mendeley Leaves and PlantVillage datasets, obtaining 91.95% and 95.80% accuracy respectively. However, the ResNet architecture was relatively computationally expensive and slow to train. GoogLeNet obtained accuracies very close to those of ResNet with 89.35% and 94.59% achieved on the Mendeley Leaves and PlantVillage datasets respectively and could be considered a less computationally expensive alternative.
- Full Text:
- Date Issued: 2021
Mobile attendance based on face detection and recognition using OpenVINO
- Authors: Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465201 , vital:76582 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9395836"
- Description: The OpenVINO toolkit enables versatile computer vision with an Intel® Movidius ™ Neural Compute Stick 2 connected to a Raspberry Pi. This small portable platform provides new opportunities for innovative solutions in computer vision applications and beyond. This paper investigates its feasibility for mobile attendance systems for settings such as classrooms or other scenarios that require headcount or roll call. Related studies of face-based systems are explored, while the advantages of the proposed system are highlighted. Although there are some positioning constraints, the proof-of-concept system processes an approximate average of five faces per second. That means it can take attendance in a lecture room of 90 students in about 18 seconds. A recognition accuracy of 98.1% with an f-score of 96.9% was yielded on a private classroom dataset captured with a modest RPi camera. These promising results were achieved using a tiny ResNet-18 architecture, producing significantly better results than MobileNet. Furthermore, it outperformed the recognition accuracy of other ‘lightweight’ methods used in the literature that do not run off embedded devices on publicly available datasets.
- Full Text:
- Date Issued: 2021
- Authors: Brown, Dane L
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465201 , vital:76582 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9395836"
- Description: The OpenVINO toolkit enables versatile computer vision with an Intel® Movidius ™ Neural Compute Stick 2 connected to a Raspberry Pi. This small portable platform provides new opportunities for innovative solutions in computer vision applications and beyond. This paper investigates its feasibility for mobile attendance systems for settings such as classrooms or other scenarios that require headcount or roll call. Related studies of face-based systems are explored, while the advantages of the proposed system are highlighted. Although there are some positioning constraints, the proof-of-concept system processes an approximate average of five faces per second. That means it can take attendance in a lecture room of 90 students in about 18 seconds. A recognition accuracy of 98.1% with an f-score of 96.9% was yielded on a private classroom dataset captured with a modest RPi camera. These promising results were achieved using a tiny ResNet-18 architecture, producing significantly better results than MobileNet. Furthermore, it outperformed the recognition accuracy of other ‘lightweight’ methods used in the literature that do not run off embedded devices on publicly available datasets.
- Full Text:
- Date Issued: 2021
Using Technology to Teach a New Generation
- Connan, James, Brown, Dane L, Watkins, Caroline
- Authors: Connan, James , Brown, Dane L , Watkins, Caroline
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465223 , vital:76584 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-030-92858-2_8"
- Description: Introductory programming courses attract students from diverse backgrounds in terms of ability, motivation and experience. This paper introduces two technological tools, Thonny and Runestone Academy, that can be used to enhance introductory courses. These tools enable instructors to track the progress of individual students. This allows for the early identification of students that are not keeping up with the course and allows for early intervention in such cases. Overall this leads to a better course with higher throughput and better student retention.
- Full Text:
- Date Issued: 2021
- Authors: Connan, James , Brown, Dane L , Watkins, Caroline
- Date: 2021
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465223 , vital:76584 , xlink:href="https://link.springer.com/chapter/10.1007/978-3-030-92858-2_8"
- Description: Introductory programming courses attract students from diverse backgrounds in terms of ability, motivation and experience. This paper introduces two technological tools, Thonny and Runestone Academy, that can be used to enhance introductory courses. These tools enable instructors to track the progress of individual students. This allows for the early identification of students that are not keeping up with the course and allows for early intervention in such cases. Overall this leads to a better course with higher throughput and better student retention.
- Full Text:
- Date Issued: 2021
Improved palmprint segmentation for robust identification and verification
- Brown, Dane L, Bradshaw, Karen L
- Authors: Brown, Dane L , Bradshaw, Karen L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/460576 , vital:75966 , xlink:href="https://doi.org/10.1109/SITIS.2019.00013"
- Description: This paper introduces an improved approach to palmprint segmentation. The approach enables both contact and contactless palmprints to be segmented regardless of constraining finger positions or whether fingers are even depicted within the image. It is compared with related systems, as well as more comprehensive identification tests, that show consistent results across other datasets. Experiments include contact and contactless palmprint images. The proposed system achieves highly accurate classification results, and highlights the importance of effective image segmentation. The proposed system is practical as it is effective with small or large amounts of training data.
- Full Text:
- Date Issued: 2019
- Authors: Brown, Dane L , Bradshaw, Karen L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/460576 , vital:75966 , xlink:href="https://doi.org/10.1109/SITIS.2019.00013"
- Description: This paper introduces an improved approach to palmprint segmentation. The approach enables both contact and contactless palmprints to be segmented regardless of constraining finger positions or whether fingers are even depicted within the image. It is compared with related systems, as well as more comprehensive identification tests, that show consistent results across other datasets. Experiments include contact and contactless palmprint images. The proposed system achieves highly accurate classification results, and highlights the importance of effective image segmentation. The proposed system is practical as it is effective with small or large amounts of training data.
- Full Text:
- Date Issued: 2019
Multi-angled face segmentation and identification using limited data
- Authors: Brown, Dane L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465711 , vital:76634 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9067899"
- Description: This paper introduces a different approach to face segmentation that aims to improve face recognition when given large pose angles and limited training data. Face segmentation is achieved by extracting landmarks which are manipulated in such a way as to normalize unseen data with a classification model. The approach is compared with related systems, followed by further tests that show consistent results across other datasets. Experiments include frontal and non-frontal training images for classification of various face pose angles. The proposed system is a promising contribution, and especially shows the importance of face segmentation. The results are achieved using minimal training data, such that both accurate and practical face recognition systems can be constructed.
- Full Text:
- Date Issued: 2019
- Authors: Brown, Dane L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465711 , vital:76634 , xlink:href="https://ieeexplore.ieee.org/abstract/document/9067899"
- Description: This paper introduces a different approach to face segmentation that aims to improve face recognition when given large pose angles and limited training data. Face segmentation is achieved by extracting landmarks which are manipulated in such a way as to normalize unseen data with a classification model. The approach is compared with related systems, followed by further tests that show consistent results across other datasets. Experiments include frontal and non-frontal training images for classification of various face pose angles. The proposed system is a promising contribution, and especially shows the importance of face segmentation. The results are achieved using minimal training data, such that both accurate and practical face recognition systems can be constructed.
- Full Text:
- Date Issued: 2019
Plant disease detection and classification for farmers and everyday gardeners
- Poole, Louise C, Brown, Dane L
- Authors: Poole, Louise C , Brown, Dane L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465722 , vital:76635 , xlink:href="https://www.researchgate.net/profile/Dane-Brown-2/publication/335378684_Plant_Disease_Detection_and_Classification_for_Farmers_and_Everyday_Gardeners/links/5d611905299bf1f70b090b54/Plant-Disease-Detection-and-Classification-for-Farmers-and-Everyday-Gardeners.pdf"
- Description: Identifying and rating diseases by hand is an expensive, time consuming, subjective and unreliable method as compared to what computers can provide. Image processing and machine learning enable automated disease identification. Research has proven that automated disease identification systems can be used as a preventative measure against plant rot and death. This paper narrows down the best techniques to segment images of leaves toward improved classification of diseases found on those leaves. An investigation is conducted on image segmentation and machine learning techniques, including state-of-the-art systems, to determine the most appropriate approach to prevent death and rot in plants. Promising results were observed during testing, and show that a system can be implemented to assist with plant health that is relevant to both home gardeners and farmers.
- Full Text:
- Date Issued: 2019
- Authors: Poole, Louise C , Brown, Dane L
- Date: 2019
- Subjects: To be catalogued
- Language: English
- Type: text , article
- Identifier: http://hdl.handle.net/10962/465722 , vital:76635 , xlink:href="https://www.researchgate.net/profile/Dane-Brown-2/publication/335378684_Plant_Disease_Detection_and_Classification_for_Farmers_and_Everyday_Gardeners/links/5d611905299bf1f70b090b54/Plant-Disease-Detection-and-Classification-for-Farmers-and-Everyday-Gardeners.pdf"
- Description: Identifying and rating diseases by hand is an expensive, time consuming, subjective and unreliable method as compared to what computers can provide. Image processing and machine learning enable automated disease identification. Research has proven that automated disease identification systems can be used as a preventative measure against plant rot and death. This paper narrows down the best techniques to segment images of leaves toward improved classification of diseases found on those leaves. An investigation is conducted on image segmentation and machine learning techniques, including state-of-the-art systems, to determine the most appropriate approach to prevent death and rot in plants. Promising results were observed during testing, and show that a system can be implemented to assist with plant health that is relevant to both home gardeners and farmers.
- Full Text:
- Date Issued: 2019