Defect Detection of Subway Tunnels Using Advanced U-Net Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Analysis

Background:

• Defect detection in real-world data presents challenges such as background-foreground imbalance, multi-scale targets, and feature similarity.
• Conventional convolutional neural network (CNN)-based methods are often insensitive to these specific issues common in semantic segmentation tasks.

Purpose of the Study:

• To propose a novel defect detection model with an improved U-Net architecture.
• To enhance the capability of detecting various types of defects, especially multi-scale targets.
• To overcome the limitations of general CNNs in handling imbalanced data and feature similarity.

Main Methods:

• The proposed method utilizes an improved U-Net architecture incorporating an atrous spatial pyramid pooling (ASPP) module and an inception module for multi-scale segmentation.
• The model is designed to specifically address the challenges of background-foreground imbalance and feature similarity.

Main Results:

• Experiments on a real-world subway tunnel image dataset demonstrated superior performance compared to general semantic segmentation and state-of-the-art methods.
• The proposed method achieved excellent detection balance across multi-scale defects.
• The model effectively identified various defect types, outperforming conventional CNN-based approaches.

Conclusions:

• The improved U-Net architecture offers a robust solution for defect detection in challenging real-world scenarios.
• The integration of ASPP and inception modules enhances multi-scale segmentation capabilities.
• This approach provides a significant advancement in semantic segmentation for infrastructure inspection and maintenance.