CNN-based model for automated anomalies recognition in facade execution to support Quality Management

Resumen

Drones are frequently used to visually inspect building envelopes to detect unsafe conditions or potential damage. Computer vision and Machine Learning (ML) technologies have emerged as promising solutions to automate image-based inspection. However, building anomalies and damages still appear during the early stages of construction due to the lack of an efficient Quality Management System (QMS). To address this gap, this study proposes an automated Convolutional Neural Networks (CNN) model to detect and classify four types of anomalies in cast-in-place concrete façades during construction, aiming to accelerate decision-making in the QMS. The research strategy adopted was the Case Study, in which nine models were created using AlexNet and ResNet architectures for training and testing with images of cast-in-place concrete wall facades collected by drones during construction. The model achieved 51.80% precision, 68.50% recall, 65.00% mAP, and an F1 score of 58.99% during training, making it the most accurate model during testing and the best among the nine. Future studies will focus on fully integrating the proposed method's workflow to enable automated analysis of drones-collected images using ML algorithms, with the automatic generation of reports based on these analyses.