Algorithms for identifying defects and inconsistencies during the acceptance of finishing works based on machine learning and computer vision in the tasks of forensic construction and technical expertise

Abstract

The article develops and validates machine learning and computer vision algorithms for objective identification of defects and nonconformities at the acceptance stage of finishing works in the context of forensic construction-technical examination; the aim is to create a software–algorithmic suite capable of increasing the reliability of conclusions, reducing expert labor intensity, and converting qualitative judgments into reproducible quantitative metrics. Materials and methods include assembling a dataset of 68,450 high-resolution images collected at 275 sites, with multi-stage expert annotation across 18 defect classes(classification, bounding-box detection, semantic segmentation with masks) and augmentation; YOLOv7 was used for detection, U-Net with an EfficientNet-B5 encoder for segmentation; training was performed on a GPU cluster with Bayesian hyperparameter optimization; evaluation on a 20% held-out set employed Accuracy, Precision, Recall, F1-score, and IoU, and practical effectiveness was compared with the results of 10 certified experts on 15 control sites. The results show a weighted-average identification accuracy of 97,84% versus 95,05% for experts, with the greatest advantage for microcracks and texture defects; segmentation achieved high IoU for linear and sharply delineated defects(over 94%), with the expected decline for amorphous stains and efflorescence, reflecting the limits of intersubjective labeling. Time costs per typical site are reduced from 19,5 person-hours to 3,45 machine-hours(-82,3%), primarily due to automated identification and measurements/quantity takeoffs, while ROC-based threshold tuning enables balancing Precision/Recall given the legal salience of defect classes. The discussion underscores the robust advantage of the integrated approach: reduced subjectivity, improved reproducibility, and scalability of documentation workflows; methodologically, integration with BIM is substantiated to reconcile as-built conditions with design, with prospects for extension to mobile/UAV data collection. The conclusion records the system’s practical applicability as a tool for standardizing and accelerating examinations, recommended for incorporation into acceptance procedures and dispute resolution.