Research on Synthetic Data Methods and Detection Models for Micro-Cracks.

Area of Science:

• Civil Engineering
• Computer Vision
• Materials Science

Background:

• Micro-crack detection in concrete is hindered by scarce labeled data, weak crack features, and complex backgrounds.
• Existing methods struggle with robustness and data availability for practical applications.

Purpose of the Study:

• To enhance micro-crack detection data availability and improve detection robustness in complex concrete surfaces.
• To develop a real-time detection system for practical infrastructure inspection.

Main Methods:

• A Poisson image editing-based synthesis strategy was used to generate realistic micro-crack training data.
• A Complex-Scene-Tolerant YOLO (CST-YOLO) detector, based on YOLOv10, was proposed with specialized modules.
• Key modules include Lighting-Adaptive Preprocessing (LAPM), Spatial-Channel Sparse Transformer (SCS-Former), and Small Object Focus Block (SOFB).

Main Results:

• The CST-YOLO detector achieved high performance with 0.990 mAP@0.5 and 0.926 mAP@0.5:0.95 at 139 FPS.
• Ablation studies confirmed the significant contributions of LAPM, SCS-Former, and SOFB modules.
• The model demonstrated effectiveness using synthesized data for training and real images for validation/testing.

Conclusions:

• Combining realistic data synthesis with an robust deep learning architecture significantly improves micro-crack detection.
• The proposed CST-YOLO detector offers a promising solution for real-time, accurate micro-crack inspection in complex environments.
• The study highlights the importance of addressing data scarcity and detection challenges in concrete structural health monitoring.