IMAGE-BASED FAçADE DEFECTS CLASSIFICATION FOR PUBLIC RESIDENTIAL BUILDINGS IN SINGAPORE � A SEMI-SUPERVISED DEEP LEARNING RESEARCH

Abstract

Residential buildings are an essential need for most people to provide a comfortable living space to call their own. For countries with a high population number yet limited land space such as Singapore, the design and construction of buildings have been increasing in both height and quantity throughout the years to cater to the demand. As buildings are subjected to constant harsh weather conditions throughout its life cycle, defects are an inevitable occurrence which not only affect the buildings’ aesthetics but also its structural integrity. With an astonishing number of residential buildings requiring timely inspections of its façade, the most efficient and safe method would be to use autonomous means via smart technologies such as drones to inspect the façade and classify the defects for swift reparations to be done. Various studies have been conducted to develop image classifiers aimed to work with defects, many of which incorporate deep learning technologies whose efficacies have been proven well. However, such classifiers have mostly been aimed at targeting defects in the civil infrastructure industry such as bridges and roads, resulting in a lack of research involving building defects for the built environment sector.

This study would focus on the usage and efficacy of deep learning technologies on the classification of building façade defects images. Traditionally, an extensive amount of data is required to achieve a high quality and performance output for the classifier, with emphasis on good quality labelled data. However, the collection of such raw images and the need for labelling by professionals is proven to be a difficult and inefficient feat. As such, the usage of a small number of labelled data with a large number of unlabelled data for classification by the means of a mean teacher algorithm will be explored in this study in order to compare the efficacy and accuracy with a traditional supervised learning approach. Based on the study in this paper, the proposed mean teacher method had achieved a prediction accuracy of 83.18% as compared to the 79.26% achieved by the supervised learning method. Moreover, the proposed mean teacher method could achieve the same accuracy as the supervised learning method using a much smaller labelled dataset. Therefore, it can be concluded that the mean teacher method is not only able to achieve better performance than the traditional method, but can also reduce the resources required for labelling of images. The incorporation of the mean teacher method in smart technologies can therefore result in a safer, more efficient and more cost-effective way of façade inspections.