CNN Training with Twenty Samples for Crack Detection via Data Augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Deep learning models, like CNNs, require extensive data for optimal performance, which is often scarce in industrial applications.
• Data augmentation is a key strategy to overcome data limitations and improve model generalization.

Purpose of the Study:

• To develop a robust crack detection system using CNNs with a limited dataset (20 images).
• To investigate and optimize data augmentation techniques for enhancing crack detection accuracy and recall.

Main Methods:

• A two-stage data augmentation approach was implemented, comparing nine augmentation methods for CNN model training.
• In-depth analysis of the rotation augmentation method and its application during inference to boost recall.
• Utilized a greedy algorithm to find optimal data augmentation combinations under resource constraints.

Main Results:

• The rotation augmentation method demonstrated superior performance compared to other techniques.
• Applying data augmentation during inference improved the detection recall of trained models.
• The proposed two-stage data augmentation significantly improved crack detection performance on a small dataset.

Conclusions:

• The developed two-stage data augmentation method effectively enhances CNN-based crack detection, even with minimal training data.
• Achieved a recall of 96% and an Fext(0.8) score of 91.18% for crack detection using only 20 training images.