Modification and Evaluation of Attention-Based Deep Neural Network for Structural Crack Detection

Hangming Yuan, Tao Jin, Xiao‐Wei YeSensors 2023 · DOI

TL;DR

This paper enhances the U-Net deep learning model with an Efficient Channel Attention (ECA) mechanism to improve structural crack detection. The proposed ECA-UNet demonstrates superior performance in identifying cracks in real-world scenarios, achieving higher recall and Intersection over Union (IOU) scores while being more robust to noise compared to baseline models.

What Problem Does This Paper Address?

Ensuring the structural integrity of buildings and bridges is paramount, and cracks are critical indicators of damage. Traditional methods for crack detection, such as visual inspections and specialized vehicles, are often inefficient, costly, and can pose safety risks. While deep learning has shown promise in automating this process, accurately identifying cracks, especially in the presence of various noise elements (like water stains, concrete stripes, or pits), remains a challenge. Existing deep learning models can struggle with distinguishing subtle crack features from background noise, leading to false positives or missed detections.

Methods

The researchers propose an improved crack detection model by integrating an Efficient Channel Attention (ECA) mechanism into the U-Net architecture.

  • Baseline Models: They first compared two established semantic segmentation networks: lraspp and U-Net.
  • U-Net Architecture: U-Net is a well-known encoder-decoder network effective for image segmentation, characterized by its skip connections that help preserve spatial information.
  • ECA Mechanism: The ECA module is a lightweight attention mechanism that focuses on channel interdependencies. It adaptively recalibrates channel-wise feature responses, allowing the network to better capture relevant information without significant computational overhead.
  • Integration: The ECA module was integrated into the upsampling path of the U-Net architecture. This placement was chosen to enhance feature extraction without interfering with the core downsampling feature extraction process.
  • Training and Evaluation: The models were trained and evaluated using standard metrics like precision, recall, and Intersection over Union (IOU).

Datasets

  • Public Dataset: The study utilized a public dataset of bridge cracks for initial training and comparison of baseline models. This dataset contained 5000 images with a 4:1 crack-to-non-crack ratio and a validation set of 1000 crack images. The crack images were of size 256x256 pixels and included various orientations.
  • Real-World Dataset: For field testing, the authors used actual structural crack images captured from on-site bridge inspections. These images were intentionally chosen to include various noise motifs (water stains, spots, joint lines, concrete stripes, scratches, pits) alongside the cracks, presenting a more challenging real-world scenario.

Key Insights and Ideas

  • Attention for Detail: The core idea is that attention mechanisms, mimicking human visual focus, can guide deep learning models to prioritize crucial crack features while downplaying irrelevant noise.
  • ECA's Efficiency: The ECA mechanism is particularly well-suited for integration with U-Net due to its lightweight nature and ability to capture channel interactions effectively without adding substantial computational burden.
  • Trade-offs in Metrics: The study highlights the inherent trade-off between precision and recall. While the ECA-UNet improved recall significantly, it saw a slight decrease in precision, underscoring the need for a balanced approach in crack detection.
  • Real-world Robustness: The most significant insight is the ECA-UNet's improved ability to handle noisy real-world images, demonstrating a more robust performance than the baseline models.

Key Results

  • Baseline Comparison: U-Net outperformed lraspp in precision (0.921 vs. 0.810), IOU (0.676 vs. 0.579), and recall (0.639 vs. 0.588) on the public dataset.
  • ECA-UNet Improvements: The ECA-UNet showed an increase of 0.016 in IOU and 0.131 in recall compared to the original U-Net.
  • Precision vs. Recall: While ECA-UNet achieved a higher recall (0.770) than U-Net (0.639), its precision was slightly lower (0.872 vs. 0.921).
  • Real-world Performance: In field tests with noisy images, lraspp recognized very few cracks. U-Net successfully detected most cracks but was susceptible to misidentifying noise motifs. The ECA-UNet accurately identified crack areas and was significantly more robust to various noise types, with only minor errors on some linear geometric noise.
  • Speed: The testing time for both U-Net and ECA-UNet was 0.058s per image, faster than lraspp (0.062s).

Limitations and Future Work

  • Noise Sensitivity: Despite improvements, the ECA-UNet can still be "cheated" by noise motifs with linear geometry.
  • Model Size: The proposed network is relatively large in terms of training parameters, suggesting a need for methods to reduce its size while maintaining performance.
  • Future Directions:
    • Further research is needed to improve robustness against crack-like linear noise motifs.
  • Investigating methods to reduce the network's size for efficiency.
  • Exploring the embedding of these models into mobile devices for real-time crack identification applications.

Summary updated

Abstract

Cracks are one of the safety-evaluation indicators for structures, providing a maintenance basis for the health and safety of structures in service. Most structural inspections rely on visual observation, while bridges rely on traditional methods such as bridge inspection vehicles, which are inefficient and pose safety risks. To alleviate the problem of low efficiency and the high cost of structural health monitoring, deep learning, as a new technology, is increasingly being applied to crack detection and recognition. Focusing on this, the current paper proposes an improved model based on the attention mechanism and the U-Net network for crack-identification research. First, the training results of the two original models, U-Net and lrassp, were compared in the experiment. The results showed that U-Net performed better than lrassp according to various indicators. Therefore, we improved the U-Net network with the attention mechanism. After experimenting with the improved network, we found that the proposed ECA-UNet network increased the Intersection over Union (IOU) and recall indicators compared to the original U-Net network by 0.016 and 0.131, respectively. In practical large-scale structural crack recognition, the proposed model had better recognition performance than the other two models, with almost no errors in identifying noise under the premise of accurately identifying cracks, demonstrating a stronger capacity for crack recognition.
BibTeX
@inproceedings{yuan2023modification,
  author = {Hangming Yuan and Tao Jin and Xiao‐Wei Ye},
  title = {{Modification and Evaluation of Attention-Based Deep Neural Network for Structural Crack Detection}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23146295},
}
Figure