Deep Crack Detection on Ceramic Material

In this paper, we focus on the problem of accurately segmenting pixels that belong to cracks and those that do not. Additionally, the cracks specifically considered in this research are small cracks on ceramic surfaces. These cracks are in the range of nanometer that need to be detected automatically before the material fractures due to crack propagation. To tackle this, we employ traditional image processing and deep learning for crack detection and segmentation. Moreover, we compare and evaluate CNN-based segmentation models with encoder-decoder architecture: U-Net, DeepLab and DeepCrack, both with and without image processing, using a dataset collected by our research team. The results show that combining traditional image processing with deep segmentation DeepLabv3+ improve the performance significantly.