Automated defect detection and classification in ashlar masonry walls using machine learning

Methods employed for surveying buildings for condition have traditionally been reliant upon visual assessment and manual recording. Survey of traditional masonry also ostensibly conforms to this approach but, due to the sheer volume of masonry units composing walls, it is often prohibitively time consuming, exceptionally complex and ultimately costly. Notable features of such survey work for ashlar stone types require each stone to be labelled and overlaid with information relative to condition. Further hindering these already costly operations, it has been shown that the accuracy of reporting, including labelling the manifestation of defects and defect diagnosis, is subjective, depending upon the expertise and experience of those evaluating the fabric. Moving beyond these preliminary survey and reporting stages, this situation gives rise to variable repair and maintenance strategies that can have significant cost implications and can debase fundamental conservation activities. The development of digital technologies, such as terrestrial laser scanning, and advancements in novel computer vision statistical techniques can help produce accurate representation of buildings that can be subsequently rapidly processed, achieving many tangible survey functions with greater inherent objectivity. In this paper, an innovative strategy for automatic detection and classification of defects in digitised ashlar masonry walling is presented. The classification method is based on the use of supervised machine learning algorithms, assisted by surveyors' strategies and expertise to identify defective individual masonry units, through to broader global patterns for groups of stones. The proposed approach has been tested on the main facade of the Chapel Royal in Stirling Castle (Scotland), demonstrating its potential for ashlar masonry forms of wall construction. It is important to recognise that the findings are not limited to this culturally significant building and will be of high value to almost innumerable ashlar-built structures worldwide. The research ultimately attempts to reduce the degree of subjectivity in classifying defects, on a scale and rapidity hitherto beyond traditional project cost constraints. Importantly, it is recognised that through automation more effective utilisation of resources that would have been traditionally spent on survey can be redeployed to support fabric intervention or routine maintenance operations.