Image-based surface scratch detection on architectural glass panels using deep learning approach

Abstract As a transparent and traditional building material, glass products such as glass facade are vital components of buildings. However, the surface scratches generated in the manufacturing process or emerging in the service stage such as windborne debris impacts may lead to remarkable strength degradation of glass material. In order to assess the fracture possibility of glass components, the size and number of scratches should be monitored during their lifecycle. Automatic scratch detection of architectural glass therefore remains a necessary task for civil engineers. A pixel-level instance segmentation model using Mask and region-based convolutional neural network (Mask R-CNN) was proposed for scratches detection on transparent glass surface. Images with scratches were firstly collected by a tailor-made automated microscopic camera scanning system to build the training and validation dataset. Test results demonstrate that the trained network is satisfactory, achieving a mean average precision of 96.5% with low missing and false rate under background interference. A comparison between the proposed model and another segmentation method YOLACT indicates that the proposed model has better performance in both detection and segmentation accuracy. The proposed deep learning-based approach can better support the development of non-contact defect assessment techniques for transparent building materials such as glass.