[Paper Review] Feature Pyramid and Hierarchical Boosting Network for Pavement Crack Detection

Fan Yang, Lei Zhang+4 more|arXiv (Cornell University)|Jan 18, 2019

Infrastructure Maintenance and MonitoringEngineering60 references61 citations

TL;DR

FPHBN introduces a feature pyramid and hierarchical boosting to improve pavement crack detection, achieving higher AIU, ODS, and OIS than several baselines on multiple datasets. It also proposes a new AIU metric for crack detection.

ABSTRACT

Pavement crack detection is a critical task for insuring road safety. Manual crack detection is extremely time-consuming. Therefore, an automatic road crack detection method is required to boost this progress. However, it remains a challenging task due to the intensity inhomogeneity of cracks and complexity of the background, e.g., the low contrast with surrounding pavements and possible shadows with similar intensity. Inspired by recent advances of deep learning in computer vision, we propose a novel network architecture, named Feature Pyramid and Hierarchical Boosting Network (FPHBN), for pavement crack detection. The proposed network integrates semantic information to low-level features for crack detection in a feature pyramid way. And, it balances the contribution of both easy and hard samples to loss by nested sample reweighting in a hierarchical way. To demonstrate the superiority and generality of the proposed method, we evaluate the proposed method on five crack datasets and compare it with state-of-the-art crack detection, edge detection, semantic segmentation methods. Extensive experiments show that the proposed method outperforms these state-of-the-art methods in terms of accuracy and generality.

Motivation & Objective

• Motivate automatic pavement crack detection to replace time-consuming manual inspection.
• Develop a deep learning architecture that fuses multi-scale context via a feature pyramid.
• Balance easy and hard samples during training with hierarchical boosting.
• Provide a robust evaluation metric (AIU) that accounts for crack width and annotation bias.
• Demonstrate generalization across multiple crack datasets.

Proposed method

• Adopt a bottom-up VGG-based feature extractor to form multi-level feature maps.
• Build a top-down feature pyramid to inject context into lower-level features.
• Attach side networks for deep supervision at each level.
• Apply hierarchical boosting to reweight samples across side networks in a top-down manner.
• Fuse multi-scale side predictions and optimize with a sigmoid cross-entropy loss.
• Evaluate using AIU, alongside traditional PR-derived metrics (ODS/OIS).

Experimental results

Research questions

• RQ1 Can a feature pyramid improve low-level crack detection by injecting high-level context?
• RQ2 Does hierarchical boosting help the model focus on hard crack/non-crack samples during training?
• RQ3 Is AIU a robust complementary metric for crack width-aware evaluation compared to PR-based measures?
• RQ4 Do the proposed components improve generalization across diverse pavement crack datasets?

Key findings

• FPHBN achieves AIU of 0.560 on CRACK500 validation, higher than HED-FP (0.553) and baseline HED (0.541).
• On CRACK500 test, FPHBN attains AIU 0.489, with ODS 0.604 and OIS 0.635, outperforming HED, RCF, and FCN baselines in these metrics.
• In ablation, adding feature pyramid to HED (HED-FP) raises AIU to 0.553, and adding hierarchical boosting (FPHBN) further raises AIU to 0.560.
• On GAPs384, FPHBN achieves AIU 0.081 with ODS 0.220 and OIS 0.231, indicating better performance than HED/RCF/FCN baselines in this dataset.
• Across evaluated methods, FPHBN shows clearer crack detections and fewer false positives compared to edge/segmentation baselines (visual results and curves).
• FPHBN runs at approximately 0.197 seconds per image on GPU for CRACK500 inference, demonstrating practical efficiency.