ICDW-YOLO: An Efficient Timber Construction Crack Detection Algorithm

Area of Science:

• Computer Vision
• Materials Science
• Structural Health Monitoring

Background:

• Accurate crack detection in wood is crucial for safety and quality control.
• Existing algorithms struggle with scale variations and data quality in wood crack detection.

Purpose of the Study:

• To develop an improved crack detection algorithm for wooden materials.
• To enhance the YOLOv8 model for better recognition of small and varied cracks.

Main Methods:

• Proposed ICDW-YOLO (improved crack detection for wooden material-YOLO) based on YOLOv8.
• Introduced novel neck network, layer structure (GSConv, GS bottleneck), and anchor algorithm with dual-layer attention.
• Implemented a gather-distribute mechanism and a higher-resolution input layer.

Main Results:

• ICDW-YOLO achieved a mAP50-95 of 79.018% on a custom wood crack dataset, a 1.869% improvement over YOLOv8.
• Demonstrated robust generalization on fire/smoke, aerial remote sensing, and coco128 datasets.
• Achieved competitive mAP50 (69.226%) and mAP50-95 (44.210%) on diverse datasets.

Conclusions:

• The proposed ICDW-YOLO algorithm effectively detects cracks in wooden materials with improved accuracy and small target sensitivity.
• The enhancements offer better feature fusion and recognition without significant complexity increase.
• ICDW-YOLO shows strong generalization capabilities across various detection tasks.