Modeling Multi-Granularity Context Information Flow for Pavement Crack Detection

Junbiao Pang, Baocheng Xiong, Jiaqi Wu, Qingming HuangIEEE Trans. ITS 2025 · DOI

TL;DR

This paper introduces MGCrackNet, a novel deep learning architecture for pavement crack detection that effectively models multi-granularity context information. By fusing local and semantic features through a context guidance module and employing a Multiple Instance Learning strategy for label alignment, MGCrackNet achieves state-of-the-art performance on challenging datasets, including a newly released, large-scale bitumen pavement crack dataset.

What Problem Does This Paper Address?

Detecting pavement cracks is a critical but challenging task for road safety and maintenance. Pavement cracks exhibit complex spatial structures, low contrast with their background, and weak spatial continuity, making them difficult to accurately identify and localize. Existing methods like semantic segmentation and object detection have limitations: semantic segmentation requires laborious pixel-level annotations, while object detection's bounding boxes fail to capture the fine details of cracks. Patch-level annotation offers a more practical alternative but introduces more noise and challenges for standard segmentation approaches. This paper aims to bridge this gap by developing a method that can effectively leverage multi-granularity context information from patch-level annotations to achieve precise crack detection.

Methods

The core of the proposed method is the MGCrackNet architecture, which integrates several key components:

  • Parallel Dilated Convolution (PDC) Module: This module acts as the backbone for extracting local context. It uses dilated convolutions with different rates (r ∈ {1, 2, 5}) to capture a larger receptive field and more contextual information than standard convolutions, while concatenating features from these dilated layers helps compensate for potential information loss.
  • Context Guidance (CG) Module: This module addresses the challenge of fusing local and semantic context. It leverages high-level semantic features from deeper network stages to guide the extraction of local features in shallower stages. A heatmap generated from deeper features acts as an attention mechanism, assigning higher weights to crack-relevant regions in the shallower features. This process is iteratively applied across multiple stages.
  • Multiple Instance Learning (MIL) Strategy: A significant challenge in patch-level annotation is the label alignment problem, where coarse patch labels cannot directly inform fine-grained pixel-level predictions. MGCrackNet employs MIL to address this. It treats each patch as a "bag" and sub-patches within it as "instances," allowing the network to learn from the noisy patch-level labels by classifying instances and then aggregating their predictions to match the patch label. This strategy is crucial for aligning context across different stages.
  • Stage-wise Training: The network is divided into five stages, and the CG module is applied iteratively. A "freeze-train" strategy is used: initially, deeper layers are trained while shallower layers are frozen, allowing semantic context to guide local feature extraction. Subsequently, shallower layers are unfrozen and optimized, enabling them to refine features based on the guidance from deeper layers.

Datasets

The paper evaluates MGCrackNet on three datasets:

  • CFD (Cement Road Crack): A public dataset with 118 crack images.
  • Cracktree: Another public dataset with 206 crack images.
  • BPC (Bitumen Pavement Crack): A newly released dataset by the authors, comprising 3,522 images. BPC is highlighted as the largest, most complex, and challenging dataset, specifically designed to address the difficulties of bitumen pavement cracks, which exhibit high similarity between cracks and background, and are prone to noise. The BPC dataset uses patch-level annotations.

Key Insights and Ideas

  • Multi-Granularity Context is Crucial: The paper strongly advocates for the fusion of both fine-grained local context (texture, edges) and coarse-grained semantic context (overall crack patterns) for accurate crack detection.
  • Semantic Guidance for Local Features: The idea of using high-level semantic information to guide the extraction of low-level local features is a key innovation. This "bootstrap-like" procedure helps focus on relevant crack regions.
  • Patch-Level Annotation Challenges: The paper clearly articulates the difficulties introduced by patch-level annotations, particularly the label alignment problem, and proposes MIL as an effective solution.
  • Bitumen Pavement Complexity: The introduction of the BPC dataset highlights the unique challenges posed by bitumen pavements, where cracks are more dynamic and harder to distinguish from the background compared to cement pavements.

Key Results

MGCrackNet demonstrates state-of-the-art performance across all three datasets.

  • On the CFD dataset, it achieves an AP of 97.29%, outperforming existing methods by a significant margin.
  • On the Cracktree dataset, it also shows superior performance, with an AP of 95.24%.
  • On the challenging BPC dataset, MGCrackNet achieves the best AP of 88.32%, significantly outperforming other methods, especially in terms of precision. While its recall on BPC is slightly lower than some methods, its higher precision indicates a better ability to distinguish cracks from noise.
  • Qualitative results show that MGCrackNet produces crack maps that are much closer to the ground truth, with fewer false positives and more continuous crack predictions compared to other methods.

Limitations and Future Work

  • Noisy Patches and Labels: The paper acknowledges that while MGCrackNet handles label alignment, dealing with inherently noisy patches and labels remains a challenge, especially for datasets like BPC.
  • Future Directions: The authors suggest future research should focus on methods to handle noisy patches and labels more effectively, as well as exploring how to utilize unlabeled data. They also propose investigating how to better handle noises and utilize unlabeled data as interesting future research directions.

Summary updated

BibTeX
@article{pang2025modeling,
  author = {Junbiao Pang and Baocheng Xiong and Jiaqi Wu and Qingming Huang},
  title = {{Modeling Multi-Granularity Context Information Flow for Pavement Crack Detection}},
  journal = {IEEE Trans. Intell. Transp. Syst.},
  year = {2025},
  doi = {10.1109/tits.2024.3438883},
}
Figure