Annotation-free defect detection for glasses based on convolutional auto-encoder with skip connections

Abstract

Glasses are fundamental materials for modern society. Therefore, strict quality control is significant for glass manufacturing. The current defect detection methods for glasses rely heavily on manual annotations. However, in some occasions, it is hard to acquire the manual annotations. To solve this problem, an unsupervised learning method, namely a convolutional auto-encoder with skip connections (SC-CAENet), is proposed in this paper. Firstly, a convolutional auto-encoder is applied to reconstruct the normal images. Afterwards, image difference is performed to detect the defect regions. In addition, skip connections and weighted structural similarity loss are introduced to improve the detection precision. The results show that the accuracy, escape rate, overkill rate, precision, recall and f1-score of SC-CAENet reach 0.9475, 0.025, 0.08, 0.924, 0.975 and 0.949, respectively. Moreover, the SC-CAENet can detect an image in 32 ms, which indicates that the SC-CAENet can precisely and efficiently detect glass defects.