Robust crack detection strategies for aerial inspection

Emanuel Aldea, Sylvie Le Hégarat‐MascleProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2015 · DOI

TL;DR

This paper investigates how well existing crack detection algorithms perform on aerial inspection images, which often suffer from motion blur. The authors found that standard methods struggle with image degradation, but propose a-contrario based strategies that show improved robustness, offering a promising direction for reliable automated infrastructure inspection.

What Problem Does This Paper Address?

Automated crack detection is crucial for efficient and safe inspection of infrastructure like roads and buildings, offering significant cost savings and security enhancements over manual methods. Unmanned Aerial Vehicles (UAVs) have become popular for this task due to their ability to access difficult areas. However, the dynamic nature of aerial imaging, including camera movement and vibrations, often leads to motion blur and other image quality issues. These degradations severely impact the performance of traditional crack detection algorithms, which are typically designed for high-quality, static images. This paper aims to bridge this gap by evaluating existing state-of-the-art crack detection techniques in the context of aerial inspection and proposing more robust alternatives.

Methods

The authors evaluate two main families of crack detection algorithms:

  • Minimal Cost Path (MCP) based strategies: Specifically, they examine the Free Form Anisotropy (FFA) algorithm. FFA assumes that a crack pixel can be characterized by a minimal path that exhibits different statistical properties compared to surrounding paths. It doesn't require training data but relies on image statistics.
  • Marked Point Process (MPP) models: MPP is a probabilistic framework that models observations using a set of points with associated "marks" (features like location, orientation, length). The authors adapt MPP models, previously used for tasks like wrinkle detection, to identify cracks by defining appropriate point characteristics and interactions.

A key contribution is the proposal of a-contrario strategies for crack detection. These methods model the "absence" of interesting structures (i.e., a defect-free scene) using a naive model. Events that are highly unlikely under this naive model are then flagged as potential defects. This approach aims to automatically determine thresholds and is less sensitive to the precise choice of parameters compared to traditional methods. The authors apply this a-contrario strategy in two stages: first, detecting "seeds" (pixels likely to be part of a fissure) based on low intensity values, and second, creating "uncommon paths" originating from these seeds.

Datasets

The paper utilizes two datasets for evaluation:

  • High-resolution dataset: A publicly available dataset (referenced as [15]) providing good quality images but lacking ground truth for defects.
  • Low-resolution dataset: A dataset from a previous study [16] with lower resolution images but containing ground truth data.

To specifically test robustness to motion blur, images from the second dataset were artificially blurred using a realistic algorithm [17] to simulate sudden motion blur.

Key Insights and Ideas

  • Image quality is paramount: The study strongly confirms that standard crack detection algorithms are highly sensitive to image quality degradation, particularly motion blur. This sensitivity is a major bottleneck for reliable aerial inspection.
  • A-contrario offers robustness: The proposed a-contrario based strategies demonstrate a significant improvement in handling motion blur. This is attributed to their ability to avoid hard thresholding and to make decisions based on a more global assessment of pixel properties relative to a background model.
  • FFA's limitations: While FFA is computationally efficient, its performance degrades drastically with motion blur. It struggles to differentiate crack paths from background noise when contrast is reduced.
  • MPP's potential: MPP models show good performance and robustness to blur, offering a more fine-grained detection than a-contrario in some cases, and highlighting local defect orientation.
  • Seed detection is critical: The initial stage of detecting potential crack pixels (seeds) is crucial for the success of subsequent path-building algorithms.

Key Results

  • The FFA algorithm showed a significant drop in performance when faced with motion blur, often failing to detect any defects.
  • Both the a-contrario strategy and MPP algorithms performed much better than FFA under motion blur conditions, exhibiting a "graceful degradation" of performance.
  • The a-contrario strategy, in particular, was highlighted for its ability to cope well with blur, as illustrated in Figure 2.
  • While FFA is computationally less complex, the proposed a-contrario and MPP methods offer superior robustness, which is essential for practical aerial inspection.

Limitations and Future Work

  • Computational Complexity: While the a-contrario strategy is robust, its computational complexity is not explicitly detailed, but it's implied to be higher than FFA.
  • Deblurring as Pre-processing: The authors suggest that deblurring could be a valuable pre-processing step to further enhance the performance of all algorithms, especially FFA.
  • Further Improvement of A-contrario: The authors believe their proposed a-contrario solutions can be further improved, potentially by incorporating more sophisticated pre-processing steps like deblurring.
  • Fusion of Strategies: The paper hints at the possibility of combining the strengths of different approaches, such as using FFA's speed for initial screening and a-contrario or MPP for more robust detection in challenging conditions.

Summary updated

Abstract

In this work, we evaluate the relevance of current state of the art algorithms widely employed in the detection of cracks, for the specific context of aerial inspection, which is characterized by image quality degradation. In this study we focus on minimal cost path and on Marked Point Process algorithms, and we test their resilience to motion blur. The results show that the current strategies for defect detection are sensitive to the quality of input images; alternatively, we suggest some improvements based on a-contrario methods that are able to cope with significant motion blur.
BibTeX
@inproceedings{aldea2015robust,
  author = {Emanuel Aldea and Sylvie Le Hégarat‐Mascle},
  title = {{Robust crack detection strategies for aerial inspection}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2015},
  doi = {10.1117/12.2182920},
}
Figure