Improving Concrete Crack Segmentation Networks through CutMix Data Synthesis and Temporal Data Fusion

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Automated visual inspection systems for concrete structures rely on accurate identification of surface defects like cracks.
• Fully convolutional neural networks (FCNs) show promise for crack segmentation but require large labeled datasets, which are often unavailable for concrete structures.
• Performance of FCNs degrades on new datasets due to variations in image conditions.

Purpose of the Study:

• To improve the accuracy and robustness of FCNs for concrete crack segmentation, especially when training data is limited.
• To address the challenge of varying image conditions and the temporal nature of crack propagation in concrete structures.

Main Methods:

• A Transfer Learning approach was developed, utilizing a U-Net pre-trained on a public dataset and fine-tuned with a synthetic dataset.
• A synthetic dataset was generated using a CutMix data augmentation technique, combining crack images with background images from target datasets.
• A novel temporal data fusion technique was proposed to aggregate predictions from multiple time steps for sequential images, enhancing crack detection recall.

Main Results:

• The proposed Transfer Learning approach improved the network's ability to distinguish cracks from background pixels.
• The temporal data fusion technique enhanced the recall of crack predictions over time.
• The combined methods resulted in a significant performance increase, with F1-score and mean Intersection over Union (mIoU) improving by 28.4% and 22.2%, respectively.

Conclusions:

• The developed Transfer Learning and temporal data fusion techniques effectively enhance FCN performance for concrete crack segmentation.
• These methods offer a robust solution for automated visual inspection systems, particularly in scenarios with limited labeled data and dynamic defect evolution.
• The study demonstrates a significant advancement in the accuracy and reliability of detecting geometric properties of concrete surface defects.