Image-Processing-Based Subway Tunnel Crack Detection System

Xiaofeng Liu, Zenglin Hong, Wei Shi, Xiaodan Guo2023 · DOI

TL;DR

This paper introduces a novel image processing and deep learning system for detecting cracks in subway tunnels. By combining advanced image enhancement techniques with a deep convolutional neural network (AlexNet), the system achieves high accuracy in identifying various crack types, offering a more efficient and reliable alternative to traditional manual inspection methods.

What Problem Does This Paper Address?

Subway tunnels, crucial infrastructure for urban transportation, are susceptible to various diseases, with cracks being a primary concern. These cracks, initially small, can propagate and compromise the tunnel's structural integrity over time due to environmental factors like temperature changes and ground settlement. Early and accurate detection of these cracks is vital for timely maintenance, preventing accidents, extending the tunnel's service life, and reducing repair costs. Traditional manual inspection methods are labor-intensive, time-consuming, prone to human error, and can be hazardous due to the tunnel environment (poor lighting, confined spaces). Existing automated methods often struggle with the complex visual conditions within tunnels, such as uneven lighting, background noise, and the subtle nature of early-stage cracks. This paper aims to bridge this gap by developing a robust and accurate automated crack detection system.

Methods

The proposed system integrates several image processing techniques with a deep learning approach.

  • Image Preprocessing:
    • Grayscale Conversion: The initial RGB images are converted to grayscale using a weighted average formula to simplify processing.
  • Erosion and Opening Operations: These morphological operations are used to filter out noise and thin crack features, enhancing contrast.
  • Linear and Nonlinear Stretching: Linear stretching adjusts the image's slope and intercept to improve contrast. Nonlinear stretching (specifically, a power transformation is mentioned) is used to expand grayscale distribution, making it more aligned with human visual perception and better highlighting crack features.
  • Image Multilevel Feature Analysis: This involves a hierarchical approach to analyze connected components within the image. Techniques like rectangularity are used to distinguish between crack-like features and noise (e.g., speckle noise, wall scratches). This step aims to filter out non-crack noise while retaining and strengthening crack features.
  • Circumscribed Rectangle Extraction: The algorithm extracts the smallest bounding rectangle around detected crack features. This process includes expanding the rectangle to better match the original image's aspect ratio, minimizing recognition errors in subsequent classification.
  • Deep Learning for Detection and Classification:
    • Single-Shot MultiBox Detector (SSD): Initially mentioned as a potential deep learning model for object detection in complex images.
  • AlexNet (Upgraded): The core of the classification and recognition is an upgraded AlexNet deep convolutional neural network. This network utilizes Batch Normalization (BN) instead of Local Response Normalization (LRN) for improved generalization and training speed. The authors highlight its suitability for extracting image features from subway tunnels.
  • Support Vector Machine (SVM): Used for comparison, particularly a cascaded SVM approach for target confirmation.

The system aims to process images, extract features, and then use deep learning models to classify and identify different types of cracks (transverse, longitudinal, oblique).

Datasets

The paper describes the creation of a sample library for training and evaluation.

  • Total Images: 9171 actual images of subway tunnels.
  • Total Target Real-Label Value Boxes: 21,547.
  • Training Set: 7130 images.
  • Validation Set: 1020 images.
  • Test Set: 1021 images.

The authors mention using a "balanced improvement method" for obtaining these datasets, suggesting an effort to ensure representative data for different crack types and conditions.

Key Insights and Ideas

  • Synergy of Image Processing and Deep Learning: The paper strongly advocates for the combined power of traditional image processing for feature enhancement and noise reduction, followed by deep learning for robust classification. This hybrid approach seems to outperform purely image-processing or simpler machine learning methods.
  • Importance of Preprocessing for Tunnel Images: The challenging visual conditions in subway tunnels (low light, complex textures) necessitate sophisticated preprocessing steps. Techniques like nonlinear stretching and multilevel feature analysis are crucial for making subtle cracks discernible.
  • AlexNet's Efficacy for Tunnel Crack Recognition: The upgraded AlexNet, with Batch Normalization, demonstrates strong performance in classifying crack types, suggesting its architecture is well-suited for the specific features of tunnel crack images.
  • Limitations of Simple Noise Reduction: While noise reduction is essential, the paper implicitly suggests that over-simplification (e.g., converting to binary images) can lead to loss of crucial crack information, impacting accuracy.

Key Results

  • Classification Accuracy:
    • Transverse cracks: 92.3%
    • Longitudinal cracks: 94.0%
    • Oblique cracks: 90.8%
    • Average accuracy for the three types: 92.4%
  • SVM vs. AlexNet: The paper states that the combined image processing and deep learning method (AlexNet) shows better performance than SVM-based classification.
  • Improved Sample Library Performance:
    • Raw Image Sample Library (AlexNet): Training 87%, Test 88.2%
    • Improved Bounding Rectangle Sample Library (AlexNet): Training 97.5%, Test 96.7%
    • This significant improvement highlights the value of the refined feature extraction and bounding box matching.
  • Detection Rates: Under a confidence threshold of 0.1, the crack detection rate reached 92.96%. At a confidence threshold of 0.5, it was 64.38%. This indicates a trade-off between detection rate and confidence, with higher confidence levels potentially missing smaller cracks.

Limitations and Future Work

  • Complicated Internal Environment: The authors acknowledge that the complex internal environment of subway tunnels, with numerous interference elements, poses challenges.
  • Limited Image Samples: The constrained number of available image samples is identified as an issue.
  • Algorithm Optimization: The algorithm still needs to be optimized and improved.
  • Future Research Directions:
    • Further fusion of image processing and deep learning techniques.
  • Improving the recognition accuracy for complex images in subway tunnels.
  • Strengthening the scientific basis and rigor of experiments.
  • Addressing issues with small cracks and the potential for repeated box selection.

Summary updated

BibTeX
@misc{liu2023image,
  author = {Xiaofeng Liu and Zenglin Hong and Wei Shi and Xiaodan Guo},
  title = {{Image-Processing-Based Subway Tunnel Crack Detection System}},
  year = {2023},
  doi = {10.3390/s23136070},
}
Figure