Defect Segmentation in Surfaces using Deep Learning

Abstract

Surface inspection is one of the most challenging tasks in the manufacturing industry. Defect classification and segmentation are the two main tasks associated with surface inspection. The major challenge lies in the collection of the dataset as it is a very costly procedure and the occurrences of defected samples are very less as compared to non defective samples. Therefore, it becomes important to devise a method that can leverage the limited data available and can also handle the class imbalance between the defected and non defected samples. In this paper, a deep learning approach is proposed that uses pertained networks to perform defect segmentation on industrial surfaces. The deep learning approach consists of an encoder and decoder architecture where on the encoder side, VGG is used with pertained imagenet weights for faster training of the model and on the decoder side, the UNet decoder model is used. The evaluation of the approach shows that the proposed method can be used for surface inspection in various industrial applications.