A YOLOv4 based defect detector for building facade inspection

Abstract

The building facade defect inspection nowadays is mainly conducted by manpower, which is costly and inefficient. It is also dangerous when the surveyors work at high levels of building. Some defects are unnoticeable by naked eyes. This may lead to potential risks to users of building. In this project, a building facade detector is implemented based on deep learning techniques. It aims to detect various categories of defects on building facade. The detector is implemented using YOLOv4 network. The well-trained detector reaches 50% overall performance on detecting 9 classes of defects and 3 other objects on building facade. The result proves the feasibility of deploying the detector model on UAV to conduct near real-time building facade defect detection, and the practicability of applying similar methods on other defect detection works.