Weakly Supervised Convolutional Neural Network for Pavement Crack Segmentation

Abstract

Crack assessment plays an important role in pavement evaluation and maintenance planning. Recent studies leverage the powerful learning capability of Artificial Neural Networks (ANNs) and have achieved good performance with computer vision-based crack detectors. Most existing models are based on the Fully Supervised Learning (FSL) approach and heavily rely on the annotation quality to achieve reasonable accuracy. The annotation cost under the FSL approach has become nontrivial and often causes heavy burdens on model development and improvement, especially for complex networks with deep layers and a large number of parameters. To combat the image annotation cost, we proposed a novel Weakly Supervised Learning U-Net (WSL U-Net) for pavement crack segmentation. With the Weakly Supervised Learning (WSL) approach, the training of the network uses weakly labeled images instead of precisely labeled images. The weakly labeled images only need rough labeling, which can significantly alleviate the labor cost and human involvement in image annotation. The experimental results from this study indicate the proposed WSL U-Net outperforms some other Semi-Supervised Learning (Semi-SL) and WSL methods and achieves comparable performance with its FSL version. The dataset cross-validation shows that WSL U-Net outperforms FSL U-Net, suggesting the proposed WSL U-Net is more robust with fewer overfitting concerns and better generalization capability.