FAçADE DEFECTS DETECTION BASED ON IMAGES USING DEEP LEARNING TECHNIQUES

Abstract

Building façade defects are inevitable as the building age and the long term exposure to harsh environmental conditions can lead to severe degradation of the building façade. Currently, Periodic Structural Inspection (PSI) is carried out by a Professional Engineer every 10 years for HDB buildings to ensure the building’s structural stability. In addition, Periodic Façade Inspection (PFI) regime has been created back in 2018 to reinforce the maintenance of building facades for buildings more than 13m tall at an interval of 7 years. Considering that 930 thousands out of a total of 1.2 million HDB units are at least 20 years old and this number is growing exponentially, there is an urgent need to automate the entire building inspection process for HDB buildings in Singapore. In the recent years, HDB has been calling for proposal with regards to automation in building inspection and the incorporation of deep learning technology for crack detection has proven to be effective in the Singapore context.

The focus of the study is to identify the effectiveness on the use of deep learning technology for image classification. The chosen technology, namely Detectron2’s Mask R-CNN and PointRend R-CNN shall demonstrate the effectiveness of image classification by identifying building façade defects based on the categorization rule. The study will evaluate the image classification model through the use of performance indicators such as Mean Average Precision (mAP) and AP at different threshold. Based on this, the results concluded that PointRend RCNN has a better performance than the widely used Mask R-CNN with the same dataset input. There were positive results from using PointRend R-CNN over the optimal Mask R-CNN algorithm in detecting different types of façade defects, with improved mAP from 21.71 to 22.10 and offers an alternative for building inspection.