Machine Learning-Assisted Vision-Based Crack Detection

Abstract

Vision-based structural health monitoring has emerged to address the rising need for automated defect detection and assessment of existing civil structures. During the last decades, the successful construction of large structures and the old age of many iconic and critical infrastructures have forced civil engineers to evolve the search for more practical and efficient methods. Despite advancements in the field with the help of vision data in conjunction with machine learning, manual methods are still predominant in the industry. While being relatively independent of object distance, orientation, and viewing angle, object recognition through visual information is seemingly effortless in humans and possesses robust adaptive generalisation. Nonetheless, manual infrastructure inspection is yet arduous, inefficient in post-disaster scenarios, and even dangerous in some cases, while the personal skills and judgement of the inspector may also influence the results. In fact, the research has shown the high accuracy of vision-based methods in controlled environments; however, uncertainties have remained regarding the robust performance of deep learning models in unprecedented situations. Aiming to overcome this issue, the low generalisation of convolutional neural networks in crack detection is studied, and Transformer-based frameworks that can achieve human-level adaptive generalisation and robustness against various noises in crack detection over different structures are proposed. Moreover, considering the lack of rich labelled datasets for supervised training and slim literature on semi-supervised methods for defect detection, a semi-supervised framework is proposed to significantly reduce the need for large annotated image datasets, without which the model may easily fail on unseen data.