Surface Scratch Detection of Monolithic Glass Panel Using Deep Learning Techniques

Abstract

Glass has been widely used in the construction sector with various kinds of applications in recent decades. However, the surface scratches generated from manufacturing process and service stage such as windborne debris impacts may lead to a strength degradation of glass material. The microscopic cracks propagation from such scratches may hence trigger glass facture unexpectedly and yield serious safety problems. In order to detect the glass damage due to such scratches, traditional manual inspection techniques have many limitations. The latest development of deep learning technology has rendered the possibility to automate such damage detection process. However, most detection methods use bounding box to roughly locate the damage in grid-cell level. To precisely describe the location of scratches, a pixel-level instance segmentation Mask R-CNN model is proposed. A total number of 1032 images with scratches are collected by a microscopic camera system to build the training and validation dataset, in which the scratches are annotated manually in pixel level. Data augmentation is adopted to improve the diversity of the dataset. During the training process, transfer learning strategy is applied to obtain the feature parameters for reducing the computation cost. Test is then performed in new architectural glass panels to evaluate the performance of the model. Test results demonstrate that the proposed trained network is satisfactory, achieving a mean average precision of 96.5% and the detection missing rate of 1.9%.