Building Surface Crack Detection and Measurement Based on Improved YOLOv11

Building surface crack detection and measurement are vital for structural health assessment. Although YOLO excels in rapid inference and generalization for object recognition, it struggles with texture-induced misdetection and missed fine cracks. This study enhances YOLOv11 by integrating CBAM to suppress background interference and implementing multiscale edge enhancement with hierarchical feature fusion for detailed crack characterization. Trained on a self-built small-sample dataset, the improved model achieves precision, mAP@50, and mAP@50-95 gains of 13%, 21% and 16.8%, respectively. When tested on both validation and unseen datasets, the enhanced model reduces false detection and missed detection rates by over 30%. Postdetection, crack skeletons and orthogonal vectors enable length/width estimation, while geometric dimensions are derived from shooting parameters. By comparing the measured values with the manual measurement results as the reference, the average relative error between the image-detected dimensions and the manually measured dimensions is within 3.5%, which verifies the engineering practicality of the method.