A deep learning approach to classifying tyres using sidewall images
- Authors: Gifford, Dean
- Date: 2019
- Subjects: Image processing -- Digital techniques , Image processing Computer science
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: http://hdl.handle.net/10948/39720 , vital:35351
- Description: End of Life Tyres (ELT's) pose a potential health and environmental risk when dumped in illegal stockpiles. For recycling to be considered feasible, a profitable business opportunity needs to be created. One method of making the recycling process of tyres more profitable is by understanding the compounds found within each tyre. This study aims at classifying these tyres in order to achieve this knowledge. A literature review was done to investigate neural networks, convolutional neural networks as well as existing deep learning architectures for image classification. A deep learning approach was applied in order to classify the logos of tyres as these approaches have proved their success in both image classification and more specifically logo classification. Although tyre classification has been implemented in the past, a deep learning approach has not been applied and the logo has not been the classifying element in any other studies. The main difference of this study compared to previous research surrounding deep learning and logo classification is the properties of the tyre logo. Logos on tyres are very similar in colour as they are purely formed in rubber and very seldom have any colour to them. Additionally, the embossed logos can contain variation among same branded tyres due to small inconsistencies in the moulds. The implementation of this deep learning solution saw multiple convolutional neural networks implemented. Some of these architectures were also implemented using transferred learning. The metrics obtained as outputs from training and testing the architectures were the accuracy, precision, recall, and F1-score. These metrics were compared in conjunction with the confusion matrix produced from testing. To ensure that variance was accounted for in the experiments, the k-fold cross-validation technique was adopted. The results of this study identified that one convolutional neural network model, MobileNet, was particularly well suited for the context of classifying logos on tyre sidewalls. The MobileNet architecture had the highest performance metrics for both training from scratch (96.7% accuracy) and transferred learning (98.8% accuracy). Three other models performed particularly well when trained from scratch, these were a modification of the LeNet architecture, ResNet50 and InceptionV3. The transferred learning results were also impressive with four out of the 5 models achieving an accuracy above 94%. Interestingly, the ResNet50 architecture failed to train when transferred learning was applied. Contrasting to this, the two models VGG16 and VGG19 failed to train when trained from scratch but performed equally as well as the other models when transferred learning was implemented. This indicates that although transferred learning can improve the performance of models, it is highly dependent on the task as well as the model. Overall the results obtained proved that a deep learning approach could be applied in order to classify tyres accurately.
- Full Text:
- Date Issued: 2019
- Authors: Gifford, Dean
- Date: 2019
- Subjects: Image processing -- Digital techniques , Image processing Computer science
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: http://hdl.handle.net/10948/39720 , vital:35351
- Description: End of Life Tyres (ELT's) pose a potential health and environmental risk when dumped in illegal stockpiles. For recycling to be considered feasible, a profitable business opportunity needs to be created. One method of making the recycling process of tyres more profitable is by understanding the compounds found within each tyre. This study aims at classifying these tyres in order to achieve this knowledge. A literature review was done to investigate neural networks, convolutional neural networks as well as existing deep learning architectures for image classification. A deep learning approach was applied in order to classify the logos of tyres as these approaches have proved their success in both image classification and more specifically logo classification. Although tyre classification has been implemented in the past, a deep learning approach has not been applied and the logo has not been the classifying element in any other studies. The main difference of this study compared to previous research surrounding deep learning and logo classification is the properties of the tyre logo. Logos on tyres are very similar in colour as they are purely formed in rubber and very seldom have any colour to them. Additionally, the embossed logos can contain variation among same branded tyres due to small inconsistencies in the moulds. The implementation of this deep learning solution saw multiple convolutional neural networks implemented. Some of these architectures were also implemented using transferred learning. The metrics obtained as outputs from training and testing the architectures were the accuracy, precision, recall, and F1-score. These metrics were compared in conjunction with the confusion matrix produced from testing. To ensure that variance was accounted for in the experiments, the k-fold cross-validation technique was adopted. The results of this study identified that one convolutional neural network model, MobileNet, was particularly well suited for the context of classifying logos on tyre sidewalls. The MobileNet architecture had the highest performance metrics for both training from scratch (96.7% accuracy) and transferred learning (98.8% accuracy). Three other models performed particularly well when trained from scratch, these were a modification of the LeNet architecture, ResNet50 and InceptionV3. The transferred learning results were also impressive with four out of the 5 models achieving an accuracy above 94%. Interestingly, the ResNet50 architecture failed to train when transferred learning was applied. Contrasting to this, the two models VGG16 and VGG19 failed to train when trained from scratch but performed equally as well as the other models when transferred learning was implemented. This indicates that although transferred learning can improve the performance of models, it is highly dependent on the task as well as the model. Overall the results obtained proved that a deep learning approach could be applied in order to classify tyres accurately.
- Full Text:
- Date Issued: 2019
Minimal motion capture with inverse kinematics for articulated human figure animation
- Authors: Casanueva, Luis
- Date: 2000
- Subjects: Virtual reality , Image processing -- Digital techniques
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:4620 , http://hdl.handle.net/10962/d1006485 , Virtual reality , Image processing -- Digital techniques
- Description: Animating an articulated figure usually requires expensive hardware in terms of motion capture equipment, processing power and rendering power. This implies a high cost system and thus eliminates the use of personal computers to drive avatars in virtual environments. We propose a system to animate an articulated human upper body in real-time, using minimal motion capture trackers to provide position and orientation for the limbs. The system has to drive an avatar in a virtual environment on a low-end computer. The cost of the motion capture equipment must be relatively low (hence the use of minimal trackers). We discuss the various types of motion capture equipment and decide to use electromagnetic trackers which are adequate for our requirements while being reasonably priced. We also discuss the use of inverse kinematics to solve for the articulated chains making up the topology of the articulated figure. Furthermore, we offer a method to describe articulated chains as well as a process to specify the reach of up to four link chains with various levels of redundancy for use in articulated figures. We then provide various types of constraints to reduce the redundancy of non-defined articulated chains, specifically for chains found in an articulated human upper body. Such methods include a way to solve for the redundancy in the orientation of the neck link, as well as three different methods to solve the redundancy of the articulated human arm. The first method involves eliminating a degree of freedom from the chain, thus reducing its redundancy. The second method calculates the elevation angle of the elbow position from the elevation angle of the hand. The third method determines the actual position of the elbow from an average of previous positions of the elbow according to the position and orientation of the hand. The previous positions of the elbow are captured during the calibration process. The redundancy of the neck is easily solved due to the small amount of redundancy in the chain. When solving the arm, the first method which should give a perfect result in theory, gives a poor result in practice due to the limitations of both the motion capture equipment and the design. The second method provides an adequate result for the position of the redundant elbow in most cases although fails in some cases. Still it benefits from a simple approach as well as very little need for calibration. The third method provides the most accurate method of the three for the position of the redundant elbow although it also fails in some cases. This method however requires a long calibration session for each user. The last two methods allow for the calibration data to be used in latter session, thus reducing considerably the calibration required. In combination with a virtual reality system, these processes allow for the real-time animation of an articulated figure to drive avatars in virtual environments or for low quality animation on a low-end computer.
- Full Text:
- Date Issued: 2000
- Authors: Casanueva, Luis
- Date: 2000
- Subjects: Virtual reality , Image processing -- Digital techniques
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:4620 , http://hdl.handle.net/10962/d1006485 , Virtual reality , Image processing -- Digital techniques
- Description: Animating an articulated figure usually requires expensive hardware in terms of motion capture equipment, processing power and rendering power. This implies a high cost system and thus eliminates the use of personal computers to drive avatars in virtual environments. We propose a system to animate an articulated human upper body in real-time, using minimal motion capture trackers to provide position and orientation for the limbs. The system has to drive an avatar in a virtual environment on a low-end computer. The cost of the motion capture equipment must be relatively low (hence the use of minimal trackers). We discuss the various types of motion capture equipment and decide to use electromagnetic trackers which are adequate for our requirements while being reasonably priced. We also discuss the use of inverse kinematics to solve for the articulated chains making up the topology of the articulated figure. Furthermore, we offer a method to describe articulated chains as well as a process to specify the reach of up to four link chains with various levels of redundancy for use in articulated figures. We then provide various types of constraints to reduce the redundancy of non-defined articulated chains, specifically for chains found in an articulated human upper body. Such methods include a way to solve for the redundancy in the orientation of the neck link, as well as three different methods to solve the redundancy of the articulated human arm. The first method involves eliminating a degree of freedom from the chain, thus reducing its redundancy. The second method calculates the elevation angle of the elbow position from the elevation angle of the hand. The third method determines the actual position of the elbow from an average of previous positions of the elbow according to the position and orientation of the hand. The previous positions of the elbow are captured during the calibration process. The redundancy of the neck is easily solved due to the small amount of redundancy in the chain. When solving the arm, the first method which should give a perfect result in theory, gives a poor result in practice due to the limitations of both the motion capture equipment and the design. The second method provides an adequate result for the position of the redundant elbow in most cases although fails in some cases. Still it benefits from a simple approach as well as very little need for calibration. The third method provides the most accurate method of the three for the position of the redundant elbow although it also fails in some cases. This method however requires a long calibration session for each user. The last two methods allow for the calibration data to be used in latter session, thus reducing considerably the calibration required. In combination with a virtual reality system, these processes allow for the real-time animation of an articulated figure to drive avatars in virtual environments or for low quality animation on a low-end computer.
- Full Text:
- Date Issued: 2000
The selection and evaluation of grey-level thresholds applied to digital images
- Authors: Brink, Anton David
- Date: 1988
- Subjects: Image processing -- Digital techniques
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5443 , http://hdl.handle.net/10962/d1001996
- Description: Many applications of image processing require the initial segmentation of the image by means of grey-level thresholding. In this thesis, the problems of automatic threshold selection and evaluation are addressed in order to find a universally applicable thresholding method. Three previously proposed threshold selection techniques are investigated, and two new methods are introduced. The results of applying these methods to several different images are evaluated using two threshold evaluation techniques, one subjective and one quantitative. It is found that no threshold selection technique is universally acceptable, as different methods work best with different images and applications
- Full Text:
- Date Issued: 1988
- Authors: Brink, Anton David
- Date: 1988
- Subjects: Image processing -- Digital techniques
- Language: English
- Type: Thesis , Masters , MSc
- Identifier: vital:5443 , http://hdl.handle.net/10962/d1001996
- Description: Many applications of image processing require the initial segmentation of the image by means of grey-level thresholding. In this thesis, the problems of automatic threshold selection and evaluation are addressed in order to find a universally applicable thresholding method. Three previously proposed threshold selection techniques are investigated, and two new methods are introduced. The results of applying these methods to several different images are evaluated using two threshold evaluation techniques, one subjective and one quantitative. It is found that no threshold selection technique is universally acceptable, as different methods work best with different images and applications
- Full Text:
- Date Issued: 1988
- «
- ‹
- 1
- ›
- »