LBA-YOLO: A novel lightweight approach for detecting micro-cracks in building structures

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Accurate building crack detection is crucial for structural integrity and safety.
• Challenges include varying crack sizes and inconsistent datasets, hindering precise localization.
• Existing methods struggle with real-time, high-accuracy detection of micro-cracks.

Purpose of the Study:

• To develop an efficient and accurate algorithm for detecting building cracks, particularly micro-cracks.
• To enhance the model's ability to handle complex backgrounds and improve semantic segmentation.
• To reduce computational overhead while maintaining high detection performance.

Main Methods:

• Utilized YOLOv8n architecture as the base model.
• Introduced AC-LayeringNetV2, a hierarchical backbone for optimized feature extraction (local, peripheral, global context).
• Incorporated RAK-Conv, a module combining attention and irregular convolution for complex background handling.

Main Results:

• Achieved a 2.20% improvement in precision and a 3.50% increase in recall.
• Demonstrated a 1.90% rise in mAP@50 compared to the baseline model.
• Reduced model size by 6.55% and computational complexity by 0.03%.

Conclusions:

• The proposed model offers practical applicability and efficiency for automatic crack detection in buildings.
• Novel integration of feature fusion and attention mechanisms effectively addresses real-time detection challenges.
• The approach shows significant potential for enhancing structural safety through advanced micro-crack identification.