Crack Detection and Classification using Digital Image Processing
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Crack detection and segmentation, and the processes of crack classification and crack index calculation which they support, are essential tools in road survey and maintenance applications. There are many image processing methods that may be applied to this domain, including traditional methods, and modern methods using machine learning. However, many challenges associated with the use of digital image processing to detect and segment cracks in images are still not solved. Some of these challenges and limitations include: (1) crack images are often noisy, have low-resolution, and contain many artifacts; (2) the associated road crack datasets are imbalanced, with only a small proportion of the data repesenting crack information; (3) three dimensional (3D) data such as crack point clouds are informative to analyze and monitor the development of crack but current acquisition methods for this data produce low-density point clouds and this problem needs to be addressed to make these data sets more useful; (4) the automated calculation of crack indices is frustrated by the lack of robust, standardised methods for automatically identifying cracks and measuring crack parameters such as crack length and crack width.
To solve the above limitations, this thesis focuses on three main contributions:
The first contribution is the proposal of a new architecture for crack detection and segmentation. This method improves the ability to segment the crack from noisy and imbalanced road crack datasets. A combination of crack detection at the region level and crack segmentation at the pixel level is shown to increase the accuracy of crack segmentation.
In the second contribution, a novel method of crack point cloud upsampling is proposed. By combining the point clouds and their corresponding 2D images in a model based on a Generative Adversarial Network framework, the proposed method aims to generate high-resolution point clouds from low-resolution point clouds and matched 2D images. The high-resolution point clouds can be used to improve the classification of crack point clouds and support crack monitoring.
The final contribution is a method to calculate crack parameters such as crack length and crack width from segmented cracks. This contribution proposes an approach for evaluating the crack length results based on the traditional metric Precision Recall Curve. The new approach is suitable for a range of narrow features such as crack lines. This thesis shows the impressive power of using digital image processing and machine learning for crack analysis in both 2D or 3D crack data.
Please use this identifier to cite or link to this item: