Autonomous Image-Based Corrosion Detection in Steel Structures Using Deep Learning.

Area of Science:

• Materials Science
• Computer Vision
• Structural Engineering

Background:

• Steel structures face environmental corrosion, posing risks of structural failure.
• Current non-destructive testing methods require manual inspection and can miss hidden defects.
• Autonomous corrosion detection is crucial for proactive maintenance and safety.

Purpose of the Study:

• To evaluate different encoder-decoder neural network architectures for automated corrosion segmentation.
• To determine the optimal training strategy for accurate detection of corroded pixels in steel structures.
• To compare the performance of various deep learning models in identifying corrosion from visual images.

Main Methods:

• Investigated encoder-decoder neural networks with DenseNet121, EfficientNetB7, ResNet34, and UNet backbones.
• Compared transfer learning versus full training strategies on segmented corroded pixels.
• Evaluated model performance using pixel-level accuracy, true-positive values, and false-positive rates on diverse datasets.

Main Results:

• EfficientNetB7 achieved the highest average pixel-level accuracy at 98.5%.
• ResNet34 demonstrated superior performance over the original UNet, especially on external datasets with varied corrosion types and colors.
• Transfer learning proved more effective than full training for deep networks, particularly with smaller datasets.

Conclusions:

• Deep learning models, particularly EfficientNetB7 and ResNet34, show significant promise for autonomous steel corrosion detection.
• ResNet34 offers a robust and accurate solution for identifying corrosion in diverse real-world scenarios.
• Optimized training strategies and network architectures are key to enhancing the reliability of automated structural health monitoring.