Open-source deep learning-based air-void detection algorithm for concrete microscopic images.

Area of Science:

• Materials Science
• Civil Engineering
• Computer Vision

Background:

• Analyzing concrete microscopic images is challenging due to heterogeneity and scale variations.
• Accurate detection of air voids is crucial for assessing concrete durability.

Purpose of the Study:

• To present an open-source deep learning algorithm for automated air-void detection in concrete microscopic images.
• To evaluate the performance of the algorithm against manual techniques and experimental data.

Main Methods:

• Development of a deep learning model using the Mask R-CNN architecture.
• Training and validation of the model on concrete microscopic images.
• Comparison of algorithm performance with manual air-void enhancement techniques.

Main Results:

• The Mask R-CNN based model achieved a mean Average Precision (mAP) of 0.6452.
• Predicted air void percentages from the model align well with experimental measurements.
• The algorithm demonstrates high precision in identifying and quantifying air voids.

Conclusions:

• The open-source deep learning algorithm offers a precise and efficient method for air-void detection in concrete.
• The model shows significant potential for future applications in evaluating concrete durability.
• Automated analysis of concrete microstructure can enhance material characterization and performance prediction.