Crack segmentation enhancement using an adaptive hyper-parameter tuning scheme configured by a novel crack type clustering method

Abstract

In the past few years, computer vision algorithms have profoundly impacted areas where image and video data are used as input data. Crack detection is one of the areas where researchers have applied countless deep learning architectures to effectively detect cracks on the surface of concrete structures and pavements. A class imbalance arises when applying crack detection at the pixel level because crack pixels occupy a small portion of the images and a small portion of the dataset. To solve this problem, researchers have proposed different approaches that most are focused on using weighted loss functions. A crack segmentation algorithm with a weighted loss function applies more weights (attention) on crack pixels during loss calculation. Even though this approach slightly improves the performance of the entire dataset, it does not include the crack type in the process of loss weighting. We explore that an adaptive weighting factor depending on the crack type results in a crack segmentation performance boost. Therefore, in this paper, we first propose an evaluation approach based on crack-type clustering. Second, an adaptive hyper-parameter tuning scheme based on the proposed clustering method is employed to improve the crack segmentation performance. Results show that applying the new hyper-parameter tuning approach improves the crack segmentation performance over the unseen data compared to the state-of-the-art approaches. Moreover, the new clustering-based evaluation approach gives in-depth insights into the performance of the crack segmentation algorithm over different crack types.