Detection and Length Measurement of Cracks Captured in Low Definitions Using Convolutional Neural Networks.

Area of Science:

• Civil Engineering
• Computer Vision
• Materials Science

Background:

• Existing crack detection methods struggle with blurry or indistinct cracks in low-resolution images.
• Previous research primarily utilized datasets with clear, well-defined crack imagery.
• There is a need for robust techniques to identify subtle cracks in real-world infrastructure.

Purpose of the Study:

• To develop and validate a framework for detecting blurred, indistinct concrete cracks.
• To evaluate the performance of Convolutional Neural Network (CNN) models on challenging crack datasets.
• To identify key factors influencing the training and effectiveness of crack detection algorithms.

Main Methods:

• A framework was proposed that segments images into small square patches for classification.
• Various established CNN models were employed to classify patches as either crack or non-crack.
• The impact of patch size and labeling strategies on training performance was investigated.
• Post-processing techniques for measuring crack lengths were developed and integrated.

Main Results:

• The proposed framework demonstrated reliable performance in detecting blurred cracks on bridge deck images.
• The study identified critical factors like patch size and labeling methods that significantly affect training outcomes.
• The performance of different CNN models was experimentally compared for crack detection tasks.
• The framework's accuracy was found to be comparable to that of human experts.

Conclusions:

• The developed framework effectively detects blurred concrete cracks, addressing limitations of previous methods.
• Optimizing patch size and labeling is crucial for successful CNN-based crack detection.
• This approach offers a viable solution for automated inspection of concrete structures with subtle defects.