Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning for Segmentation of Cracks in High-Resolution Images of Steel Bridges

Andrii Kompanets, Gautam Pai, Remco Duits, Davide Leonetti, Bert Snijder

TL;DR

This work tackles automatic segmentation of fatigue cracks in high-resolution steel-bridge images captured by UAVs. It introduces a ConvNext-based encoder–decoder architecture with a specialized loss-inversion scheme and deep supervision to effectively train on patch-based, high-resolution data and reduce false positives. A new public dataset (CSB) of steel-bridge cracks, plus CSB patch variants, enables thorough evaluation of training strategies, background-patch effects, and annotation impact. Results show that background-patch-aware training and loss-inversion significantly improve performance, and patch size and global context substantially influence crack discrimination, with practical implications for UAV-aided bridge inspection workflows.

Abstract

Automating the current bridge visual inspection practices using drones and image processing techniques is a prominent way to make these inspections more effective, robust, and less expensive. In this paper, we investigate the development of a novel deep-learning method for the detection of fatigue cracks in high-resolution images of steel bridges. First, we present a novel and challenging dataset comprising of images of cracks in steel bridges. Secondly, we integrate the ConvNext neural network with a previous state-of-the-art encoder-decoder network for crack segmentation. We study and report, the effects of the use of background patches on the network performance when applied to high-resolution images of cracks in steel bridges. Finally, we introduce a loss function that allows the use of more background patches for the training process, which yields a significant reduction in false positive rates.

Deep Learning for Segmentation of Cracks in High-Resolution Images of Steel Bridges

TL;DR

This work tackles automatic segmentation of fatigue cracks in high-resolution steel-bridge images captured by UAVs. It introduces a ConvNext-based encoder–decoder architecture with a specialized loss-inversion scheme and deep supervision to effectively train on patch-based, high-resolution data and reduce false positives. A new public dataset (CSB) of steel-bridge cracks, plus CSB patch variants, enables thorough evaluation of training strategies, background-patch effects, and annotation impact. Results show that background-patch-aware training and loss-inversion significantly improve performance, and patch size and global context substantially influence crack discrimination, with practical implications for UAV-aided bridge inspection workflows.

Abstract

Automating the current bridge visual inspection practices using drones and image processing techniques is a prominent way to make these inspections more effective, robust, and less expensive. In this paper, we investigate the development of a novel deep-learning method for the detection of fatigue cracks in high-resolution images of steel bridges. First, we present a novel and challenging dataset comprising of images of cracks in steel bridges. Secondly, we integrate the ConvNext neural network with a previous state-of-the-art encoder-decoder network for crack segmentation. We study and report, the effects of the use of background patches on the network performance when applied to high-resolution images of cracks in steel bridges. Finally, we introduce a loss function that allows the use of more background patches for the training process, which yields a significant reduction in false positive rates.
Paper Structure (33 sections, 14 equations, 9 figures, 10 tables)

This paper contains 33 sections, 14 equations, 9 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Deep learning for crack detection
  3. Crack Detection in Steel Bridges
  4. Contributions of the Article
  5. Structure of the Article
  6. Models and Methods
  7. Encoder
  8. Decoder
  9. Loss function with inversion
  10. Sum of Dice loss and BCE loss
  11. Loss function with inversion
  12. Deep supervision
  13. Datasets
  14. CFD
  15. Cracks in the Steel Bridges (CSB) dataset
  16. ...and 18 more sections

Figures (9)

  • Figure 1: Scheme of the proposed encoder-decoder network
  • Figure 2: Examples of images and ground truth crack segmentation from the CFD dataset
  • Figure 3: Examples of images and ground truth crack segmentation from the CSB dataset
  • Figure 4: Illustration of the datasets structure
  • Figure 5: Example of an image of size 4608x3456 pixels split into patches of size 512x512. Red squares represent background patches and green squares represent crack patches. Note: the bottom row of patches overlaps with the row above it
  • ...and 4 more figures