Policy Gradient–Based Focal Loss to Reduce False Negative Errors of Convolutional Neural Networks for Pavement Crack Segmentation

Abstract

Convolutional neural networks (CNNs) have achieved tremendous success in pavement crack segmentation. However, it is difficult for CNN-based crack segmentation methods to minimize false-negative and false-positive errors. Compared with false-positive errors, false-negative errors are more difficult to observe and reduce manually. This paper proposes a fine-tuning method for trained CNNs, called policy gradient-based focal loss (focal-PG loss). The trained CNNs will be further trained by focal-PG loss for only one epoch. The proposed focal-PG loss can be applied to reduce the false-negative errors of the trained CNNs by sacrificing their precision. The experimental results show that focal-PG loss greatly improves the crack recognition rate of the trained encoder–decoder network (EDNet). EDNet (focal-PG loss) achieves an overall precision of 96.05%, recall of 99.68%, and F1-score of 97.83% on 100 validation images. In addition, overall precision of 95.53%, recall of 99.58%, and F1-score of 97.51% are observed for the 150 testing images. U-net, LinkNet, and the feature pyramid network are also tested in the paper to validate the effectiveness of focal-PG loss. The results demonstrate that the focal-PG loss can also improve the performance of the aforementioned networks.