Towards identification of long-term building defects using transfer learning

Detecting long-term issues on various types of building wall surfaces, such as cracks, flakes, and roofs, is vital for timely maintenance and repairs before they become too risky and expensive. Currently, building managers manually assess the building conditions to survey and communicate the state of their buildings. However, this manual process is subjective, often leads to inaccuracies, and is time-consuming; thus, it needs to be more efficient. These flaws can severely influence a building's structural stability if they go undiscovered and ignored. In this context, this study proposes an approach named towards identification of long-term building defects using transfer learning (TILT) to identify unnoticed defects such as cracks, flakes, and roofs robustly and accurately in buildings. The proposed model has been tested using images taken from real-world onsite deployments, and the types of construction issues have been determined with 98.33% accuracy predicted by the VGG16 model and 79.13% accuracy predicted by the ResNet50 model. Overall, the VGG16 model gives better results compared to ResNet50.