Machine learning for automatic detection of building structure defects based on visual inspection

Abstract

Manual visual inspection of building structures remains dependent on the surveyor qualification and subjective perception, which keeps the agreement between conclusions issued for the same object by different specialists rather low. The study proposes and computes a model for the automatic detection of defects in load-bearing and enclosing structures from visual-inspection photographs, based on a comparison of four machine-learning architectures – the ResNet-50 classifier, the YOLOv8 detector and the U-Net and DeepLabv3+ segmentation networks. On a model sample of 12,480 annotated fragments covering five defect classes, precision, recall and the F1 score are obtained for every class, together with the mAP indicator and the mask intersection area. The highest recall for cracks with an opening from 0.25 mm is reached by DeepLabv3+ (0.921 against 0.864 for the detector), while precision on fine efflorescence drops to 0.68. It is established that the mismatch between automatic annotation and the manual protocol concentrates on defects with a blurred boundary and under unfavourable illumination, whereas the gain in reproducibility of the conclusion is obtained by moving the routine screening onto the model while retaining expert verification of borderline cases. It is shown that the decisive parameter of performance is the spatial resolution of the image: at a resolution coarser than 0.8 mm per pixel the difference in metrics between the networks shrinks to a practically negligible value, and the choice of architecture ceases to affect the recognition outcome more than the shooting conditions.