Abstract
To investigate the performance of convolutional neural network models in the field of building crack detection, this study selected the lightweight and efficient YOLOv11-seg model for object detection experiments on building cracks. By using polygons to annotate cracks at the pixel level, the model calculated experimental results for both bounding boxes and masks. The results show that the accuracy and recall rate of object detection both exceed 70%, indicating that the YOLOv11-seg model performs well in detecting building cracks. When the IoU threshold is 0.5, the model achieves a mean average precision (mAP) of over 75%, demonstrating its strong recognition capability for building cracks under moderate precision requirements. However, when the IoU threshold ranges from 0.5 to 0.95, the average mAP significantly decreases, and the mAP for bounding box prediction is notably higher than that for mask prediction, indicating that the model's accuracy in predicting crack edges is insufficient. These experimental results suggest that the YOLOv11-seg model can rapidly locate building cracks but still requires improvements in the precise detection of crack edges.
Introduction
Building cracks are critical indicators of structural health and load-bearing capacity. Traditionally, inspections are performed manually—a process that is slow, subjective, and potentially hazardous for high-rise structures. While computer vision has long attempted to automate this, early methods like Sobel filters often fail in real-world conditions due to complex backgrounds and varying light. Deep learning offers a solution, but a trade-off remains between the speed required for mobile inspections and the precision needed to measure crack dimensions. This study evaluates whether the YOLOv11-seg architecture can bridge this gap.
Methodology
The researchers deployed YOLOv11-seg, an instance segmentation model released in September 2024. This model evolves the YOLO architecture with several key technical upgrades designed for efficiency and small-target detection.
Key architectural features include:
- C3K2 Modules: These replace older C2F modules in the backbone to reduce computational complexity while enhancing the detection of small targets.
- C2PSA Module: Positioned after the Spatial Pyramid Pooling (SPPF) layer, this adds a parallel spatial attention mechanism to help the model focus on slender crack lines amidst noisy backgrounds.
- Depthwise Separable Convolutions: Utilized in the detection head to maintain accuracy while significantly boosting processing speed.
Experimental Setup
The study utilized a custom dataset of 800 original images of building cracks. Images were standardized to a 500x500 pixel resolution and augmented using horizontal flipping and brightness adjustments to improve robustness. The data was split into a 7:2:1 ratio for training, validation, and testing. Annotations were performed using Labelme to create polygon-based masks, allowing the model to learn pixel-level delineation.
Results and Discussion
The experimental results demonstrate that the YOLOv11-seg model is highly effective at localization but faces challenges in fine-grained delineation. Both Precision and Recall exceeded 70% for boxes and masks, with a solid [email protected] of 76.4% (Box) and 76.5% (Mask).
However, a significant "Precision Gap" emerged at higher IoU thresholds. For the [email protected]:0.95 metric, the Box score remained respectable at 60.5%, but the Mask score plummeted to 33.8%. This suggests that while the model excels at "finding" the crack (localization), it struggles to "trace" the slender, irregular edges perfectly (delineation). Bounding boxes proved more robust than masks at high precision requirements, indicating that the segmentation head requires further refinement for thin structural defects.
Conclusion
YOLOv11-seg provides a rapid and reliable means of locating building cracks, making it suitable for initial on-site surveys. However, its current edge-prediction accuracy is insufficient to replace manual micrometers for measuring exact crack widths. Future work will focus on improving edge delineation and developing automated methods to calculate physical crack dimensions directly from segmented masks.
Summary updated