A deep learning-based building defects detection tool for sustainability monitoring

Abstract

To ensure sustainability of buildings, detection of building defects is crucial. Conventional practices of defects detection from building inspection data are mostly manual and error prone. With the advancements in computer vision, imaging technology and machine learning-based tools have been developed for real-time, accurate and efficient defects detection. Deep learning (DL), which is a branch of ML is more robust in automatically retrieving elements’ semantics to detect building defects. Although DL algorithms are robust in object detection, the computational complexities and configurations of these models are high. Therefore, this study presents a process of developing a computationally inexpensive and less complicated DL model using transfer learning and Google Colab virtual machine to improve automation in building defects detection. Cracks is one of the major building defects that constraint the safety and durability of buildings thus hindering building sustainability. Building cracks images were sourced from the Internet to train the model, which was built upon You Only Look Once (YOLO) DL algorithm. To test the DL model, inspection images of five (05) buildings collected by the Facilities Management department of a University in Sydney city were used. The DL model developed using this process offers a monitoring tool to ensure the sustainability of buildings with its’ ability of detecting cracks from building inspection data in real time accurately and efficiently. Although the current model is built to detect cracks, this process can be employed to automated detection of any building defect upon providing the training images of defects.