Defect Detection of Subway Tunnels Using Advanced U-Net Network

An Wang, Ren Togo, Takahiro Ogawa, Miki Haseyama

TL;DR

This paper introduces an enhanced U-Net architecture for detecting various defects in subway tunnel images. By incorporating Inception modules and an Atrous Spatial Pyramid Pooling (ASPP) module, their model effectively addresses challenges like imbalanced data and multi-scale defects, outperforming existing methods in both quantitative and qualitative evaluations.

What Problem Does This Paper Address?

Maintaining aging infrastructure, particularly subway tunnels, is crucial for safety and economic reasons. Traditional manual inspection methods are time-consuming and labor-intensive. While AI and deep learning offer promising solutions for automated defect detection, existing methods struggle with specific challenges in subway tunnel imagery:

  • Background-Foreground Imbalance: Tunnel images often have a large background area with limited defect regions, making it difficult for models to learn effectively.
  • Multi-Scale Defects: Defects like cracks, cold joints, and water leaks vary significantly in size and appearance, requiring a model that can capture these diverse scales.
  • Complex Backgrounds: Tunnel walls can have structures that visually resemble defects, leading to false positives.

This paper aims to develop a more robust and accurate automated system for detecting these diverse defects in subway tunnels.

Methods

The authors propose an advanced U-Net architecture tailored for subway tunnel defect detection. Their approach builds upon the U-Net's strengths in semantic segmentation, particularly its ability to preserve fine details through skip connections. The key modifications include:

  • Inception Modules: Replacing standard convolutional blocks with Inception modules (inspired by Inception-v3) to enhance the network's receptive field and improve its ability to capture features at different scales. This helps in detecting multi-scale defects.
  • Atrous Spatial Pyramid Pooling (ASPP) Module: Integrating an ASPP module (from Deeplab-v3+) at the end of the encoder. This module uses atrous convolutions with different rates to capture contextual information at multiple scales without downsampling, further aiding in multi-scale defect detection. The authors adapted the atrous rates to suit the smaller input size of their subway tunnel images.
  • Data Augmentation and Patch Selection: To address the background-foreground imbalance, they employed a strategy of dividing high-resolution images into patches and experimentally balancing the ratio of defect to background patches to approximately 1:1 for training.

Datasets

The study utilized a real-world subway tunnel image dataset provided by Tokyo Metro Co., Ltd. The original images were high-resolution (approximately 12,088 × 10,000 pixels with 1 mm/pixel resolution). To handle these large images, they were divided into smaller patches of 256 × 256 pixels.

  • Training Set: 280,000 patches from 29 images, with a 1:1 ratio of background to defect patches.
  • Validation Set: 71,818 patches from 7 images.
  • Test Set: 326,172 patches from 11 images, using the same patch division strategy without discarding background patches.

Key Insights and Ideas

  • U-Net as a Strong Foundation: The U-Net architecture's inherent ability to capture both high-level semantic information and low-level spatial details makes it a suitable backbone for defect detection in images where fine features are critical.
  • Synergy of Inception and ASPP: Combining Inception modules for multi-scale feature extraction and ASPP for capturing context at different dilation rates proved effective in handling the diverse nature of subway tunnel defects.
  • Importance of Data Balancing: The significant improvement achieved by balancing background and defect patches highlights the critical impact of data distribution on the performance of segmentation models, especially in real-world scenarios with inherent imbalances.
  • Adaptation is Key: Modifying existing modules like ASPP to suit the specific characteristics (e.g., input size) of the target dataset is crucial for achieving optimal performance.

Key Results

The proposed method, referred to as PM, demonstrated superior performance compared to several baseline and state-of-the-art methods (CM1-CM7).

  • Overall Performance: PM achieved the highest IoU (0.356) among all tested methods, indicating better overall segmentation accuracy. It also showed strong performance in Recall (0.660) and F-Measure (0.525).
  • Defect-Specific Recall: In terms of recall for individual defect types, PM excelled in detecting larger defects like "Peeling" (0.921), "Patching (intermediate pile)" (0.819), and "Floating" (0.802). While it showed limitations in detecting very small cracks (e.g., 0.3-0.5 mm), the qualitative analysis suggested that these limitations might not significantly impact practical inspection tasks.
  • Qualitative Analysis: Visual results showed that PM produced fewer false detections compared to other methods, leading to more reliable outputs for inspectors.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Specificity: The study used data from a specific subway line in Japan. Future work should explore the model's generalizability to data from different subway lines and construction methods, as wall conditions can vary significantly.
  • Dataset Size: While a substantial number of patches were used, increasing the number of high-resolution images for patch extraction could further enhance the model's robustness.
  • Small Crack Detection: The model still faces challenges in detecting very small cracks, which could be an area for future improvement.

For future work, the authors plan to investigate new strategies for enhancing detection accuracy and explore the application of their model to other real-world tasks.

Summary updated

Abstract

In this paper, we present a novel defect detection model based on an improved U-Net architecture. As a semantic segmentation task, the defect detection task has the problems of background–foreground imbalance, multi-scale targets, and feature similarity between the background and defects in the real-world data. Conventionally, general convolutional neural network (CNN)-based networks mainly focus on natural image tasks, which are insensitive to the problems in our task. The proposed method has a network design for multi-scale segmentation based on the U-Net architecture including an atrous spatial pyramid pooling (ASPP) module and an inception module, and can detect various types of defects compared to conventional simple CNN-based methods. Through the experiments using a real-world subway tunnel image dataset, the proposed method showed a better performance than that of general semantic segmentation including state-of-the-art methods. Additionally, we showed that our method can achieve excellent detection balance among multi-scale defects.
BibTeX
@inproceedings{wang2022defect,
  author = {An Wang and Ren Togo and Takahiro Ogawa and Miki Haseyama},
  title = {{Defect Detection of Subway Tunnels Using Advanced U-Net Network}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=xH80WMQiUo},
}
Figure