Date

2023-1-27

Author

Muturi, Tanner Wambui

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

479
views

26
downloads

Cite This

Cracks are observed to be an initial sign of the degradation of the pavement and should therefore be detected and repaired to prevent further disintegrations. With the growth in technology, various image processing, machine learning, and deep learning methods have been applied to detect cracks on road pavements. The research leans towards using deep learning models for the pixel-wise segmentation of pavement image cracks among these models. However, most deep-learning models adopted in the literature are supervised and thus require enormous amounts of images with their corresponding ground-truth labels, which are expensive to obtain. Therefore, this study proposes using Cycle Generative Adversarial Network (CycleGAN), an unsupervised image translation model, for the pixel-wise segmentation of crack regions. A novel training procedure is adopted, with the forward and reverse cycle generators trained for every odd epoch and the discriminators for the even ones. A rigid pavement dataset of Fully Convolutional Network (FCN) (X. Yang et al., 2018) and drone (Ersoz et al., 2017) images are collected for training. As the model does not require accurate ground truth labels, labor-free unpaired image labels are obtained by compiling the ground truth crack labels from CrackForest, GAPS384, CrackTree200, and Crack500 public datasets. The proposed model achieved an overall F1 score of 0.84, achieving comparable results with the CrackForest algorithm (Shi et al., 2016), FCN algorithm (X. Yang et al., 2018), and pix2pix model (Isola et al., 2017). In addition, the model outperforms the CrackForest algorithm and FCN-supervised model when testing the drone dataset. Finally, the effect of change in discriminator architecture and the application of transfer learning in the generator is investigated. A one-class discriminator architecture and loading of ImageNet pre-trained weights to the generator were observed to achieve the best performance.

Subject Keywords

Unsupervised crack detection, Generative adversarial networks, Cycle generative adversarial networks, Image translation, Segmentation

URI

https://hdl.handle.net/11511/102524

Collections

Graduate School of Natural and Applied Sciences, Thesis