Evaluation-oriented façade defects detection using rule-based deep learning method

Abstract

Visual evaluation of façade condition plays an important role to ensure the structural health of the whole building. To automatically achieve visual evaluation of façade condition with high accuracy, current studies have applied various machine learning and deep learning algorithms to classify, localize, and segment the defects. However, the methods in previous research mainly focused on accuracy improvement rather than providing effective evaluations of defects according to the requirement of industry standards. Therefore, this study proposes a rule-based deep learning method to achieve evaluation-oriented façade defects detection, which can be used to provide effective evaluation areas containing the necessary information (e.g., type, location, quantity, and size of façade defects) for condition evaluation. First, annotation rules for classification, segmentation, and localization are designed to instruct the manual annotation work and automatically adjust the bounding boxes into effective evaluation areas. Then, a proposal weighting rule is developed to be combined with the deep learning model during the model training process to improve the accuracy and stability of the predictions. A rectification rule is further used to adjust the raw predictions into predictions with effective evaluation areas for façade defects. Experiments conducted in this study demonstrated that using the proposed method can successfully improve the performance of façade defects detection to meet the requirement of condition evaluation. Besides, this method is tested to be adaptable to various distance settings in the requirement.