A global context and pyramidal scale guided convolutional neural network for pavement crack detection

Pavement crack detection is a crucial part of road maintenance. Traditional crack detection methods are time-consuming and unreliable. Therefore, researchers have adapted deep-learning-based segmentation approaches from several computer vision applications for crack detection. However, these approaches are not always suited for small objects, such as crack segmentation, because they will miss precise crack information, which occupies only 5-15% of pixels in the whole image compared to the background pixels. To address this issue, we introduce a feature fusion module to the encoder-decoder architecture, considerably improving the ability to acquire detailed information on crack features. Two separate branches of this module are used to maintain and improve the global and multi-scale contexts of crack images. Additionally, the sum of cross-entropy, Tversky, and lovasz hinge losses is used as a loss function for the imbalanced distribution of crack and background pixels. To prove the superiority of the proposed approach, we used four public datasets. Our approach achieves precision of 0.8413, recall of 0.8120, and intersection over union (IoU) of 0.6553 on the Crack500 dataset; precision of 0.9520, recall of 0.9408, and IoU of 0.8982 on the DeepCrack dataset; precision of 0.9177, recall of 0.9148, and IoU of 0.8455 on the GAPs384 dataset; and precision of 0.8552, recall of 0.8273, and IoU of 0.6738 on the MCD dataset.