A Fast Dent Detection Method for Curved Glass Using Deep Convolutional Neural Network

Abstract

The curved glass is widely used in many fields, but its defects inspection is still a labor-intensive job. In all kinds of defects in glass, the dent defect is the hardest one because of its small depth variation and smooth edge. Machine vision gives out a possible solution for defects detection in glass industry, but the dent images suffer from the non-uniform gray value and the low contrast. In this paper, we propose a method based on the deep convolutional neural network for the dent defect detection. We prune the DenseNet-121 to design a compact model for real-time production. During the process of model training, we use a data augmentation method including offline and online operations to optimize the model performance. The experiments show this detection method has a good performance of 100% recognition accuracy on our dent defect dataset of the curved glass.