A Novel Denoising Autoencoder Method for Surface Defect Detection of Screw Products

Abstract

Defect detection is an important aspect of assessing the surface quality of screw products. A defective screw greatly affects the mechanism of screw product. Recently, unsupervised learning has been widely used for defect detection in industrial applications. In most cases, anomaly networks are unable to reconstruct abnormal images into satisfactory normal images, which results in poor defect detection performance. In this paper, a denoising autoencoder is used to enhance the capability of reconstructing defect screw images. By using this technique, the model can efficiently extract more features during reconstruction. Compared to the results without noise, the IoU can be increased by over 11%. The paper also develops an intelligent screw detection system for realistic industrial applications. Consequently, the proposed scheme is well suited to industrial defect detection scenarios since the models require only normal samples for training.