Detection of Steel Surface Defects Using U-Net with Pre-trained Encoder

Abstract

Steel is one of the most widely building materials of modern times. Automatic detection of manufacturing defects in steel surfaces is a very important step for product quality control in the steel manufacturing industry. Traditional inspection systems require high maintenance costs and allow little flexibility. In this paper, we propose using the deep learning network U-Net for detecting defects. U-Net is a type of Convolutional Neural Networks that perform semantic segmentation of images, it was originally developed to achieve biomedical image segmentation, as U-Net requires smaller training sets compared with typical CNNs. We will present that segmentation can be done with promising results when using the U-Net model with pre-trained encoder; as transfer learning increases the accuracy and stability of the object detection models.