Machine Learning for Improved Detection and Segmentation of Building Boundary
| dc.contributor.advisor | Mourshed, Monjur | |
| dc.contributor.author | Algarni, Salem | |
| dc.date.accessioned | 2024-07-11T12:35:11Z | |
| dc.date.available | 2024-07-11T12:35:11Z | |
| dc.date.issued | 2022-09-27 | |
| dc.description | The first step in rescuing and mitigating the losses from natural or man-made disasters is to assess damaged assets, including services, utilities and infrastructure, such as buildings. However, manual visual analysis of the affected buildings can be time consuming and labor intensive. Automatic detection of buildings, on the other hand, has the potential to overcome the limitations of conventional approaches. This thesis reviews the existing methods for the automated detection of objects using multi-source geospatial data and presents two novel post processing techniques. Effective building segmentation and recognition techniques are also investigated. Artificial intelligence techniques have been used to identify building boundaries in automated building-detection applications. Compared with other neural network models, the convolutional neural network (CNN) architectures based on supervised and unsupervised approaches provide better results by looking at the image details as spatial information of the entity in the frame. This research incorporates the improved semantic detection ability of Region-based Convolutional Neural Network (Mask R-CNN) and the segmentation refining capability of the conditional random field (CRF)s. Mask R-CNN uses a pre-trained network to recognize the boundary boxes around buildings. It also provides contour key points around buildings that are masked in satellite images. This thesis proposes two novel post-processing techniques that operate by modifying and detecting the building’s relative orientation properties and combining the key points predicted by the two head neural networks to modify the predicted contour with the help of the proposed novel snap algorithms. The results show significant improvements in the accuracy of boundary detection compared with the state-of-the-art techniques of 2.5%, 4.6% and 1% for F1-Score, Intersection over Union also known as Intersection over Union, and overall pixel accuracy, respectively. CNNs have proven to be powerful tools for a wide range of image processing tasks where they can be used to automatically learn mid-level and high level concepts from raw data, such as images. Finally, the results highlight the potential of further approaches to these applications, such as the planning of infrastructure. | |
| dc.description.abstract | This thesis addresses the need for rapid assessment of damaged assets, such as buildings, following natural or man-made disasters. Traditional manual visual analysis is labor-intensive and time-consuming, prompting the exploration of automated detection methods using multi-source geospatial data. The research reviews existing object detection methods and introduces two novel post-processing techniques. Artificial intelligence, particularly convolutional neural networks (CNNs), is employed to improve building detection accuracy. The study emphasizes the superior performance of CNN architectures, especially Region-based Convolutional Neural Network (Mask R-CNN), which enhances semantic detection and boundary recognition. Mask R-CNN, combined with conditional random fields (CRFs), effectively identifies and refines building contours in satellite images. The proposed post-processing techniques modify the relative orientation properties of buildings and integrate key points from two neural networks to adjust predicted contours using innovative snap algorithms. The results demonstrate notable improvements in boundary detection accuracy, with enhancements of 2.5% in F1-Score, 4.6% in Intersection over Union, and 1% in overall pixel accuracy compared to current state-of-the-art methods. The thesis highlights the potential of CNNs in automating image processing tasks, learning complex concepts from raw data, and aiding infrastructure planning and disaster response. | |
| dc.format.extent | 252 | |
| dc.identifier.uri | https://hdl.handle.net/20.500.14154/72568 | |
| dc.language.iso | en | |
| dc.publisher | Cardiff University | |
| dc.subject | Mask R-CNN | |
| dc.subject | Convolutional neural networks (CNNs) | |
| dc.subject | Image processing | |
| dc.subject | Region-based Convolutional Neural Network (Mask R-CNN) | |
| dc.subject | Artificial intelligence | |
| dc.subject | Boundary detection | |
| dc.subject | Semantic detection | |
| dc.subject | Building segmentation | |
| dc.subject | Automated building detection | |
| dc.subject | Infrastructure planning | |
| dc.subject | Multi-source geospatial data | |
| dc.title | Machine Learning for Improved Detection and Segmentation of Building Boundary | |
| dc.type | Thesis | |
| sdl.degree.department | Engineering | |
| sdl.degree.discipline | Machine Learning & AI | |
| sdl.degree.grantor | Cardiff University | |
| sdl.degree.name | Doctor of Philosophy |