Fast Detection of Missing Thin Propagating Cracks during Deep-Learning-Based Concrete Crack/Non-Crack Classification.

Area of Science:

• Civil Engineering
• Computer Science
• Materials Science

Background:

• Existing deep learning (DL) models struggle to detect thin, single-pixel width cracks.
• Accurate crack detection is crucial for concrete structure integrity assessment.

Purpose of the Study:

• To propose a computationally efficient scheme for tracking thin/propagating crack segments missed by DL models.
• To enhance DL-based crack identification on concrete surfaces.

Main Methods:

• A hybrid approach combining image processing (pre- and post-processor) with a 1D DL model.
• Image processing assists DL by identifying crack candidate regions and tracking thin cracks.
• The method is validated on low-resolution UAV-captured images with varying concrete textures and disturbances.

Main Results:

• Successfully tracks thin cracks (single-pixel width) missed by conventional DL models.
• Demonstrates robustness across diverse concrete surface textures, lighting conditions, and complex scenes.
• The approach is invariant to initial sensitivity parameters and hyperparameters due to multi-threshold image processing.

Conclusions:

• The proposed image-processing-assisted DL scheme offers an efficient and accurate alternative for thin crack detection in concrete.
• It overcomes the limitations of semantic segmentation for pixelated mapping of fine crack regimes.
• This method reduces the need for labor-intensive and skilled manual labeling in crack assessment.