An Automated Visual Defect Segmentation for Flat Steel Surface Using Deep Neural Networks

Defect detection on the metal sheets is an importance for the steel production industry. This paper proposes an automated visual defect segmentation for flat steel surface using two different deep neural network methods, namely U-Net and fully convolutional network-8 (FCN-8). Different models like ResNet34, MobilNetV2, EfficientNetB0, InceptionV3, and VGG16 have been implemented as the U-Net’s encoder. In addition, FCN8 model has been investigated as a different architecture for the SEVERESTAL dataset. Due to the imbalanced dataset, various techniques such as dice coefficient as loss function and data augmentation have been used. Moreover, no-defect steel surfaces have been considered as a separate class. As a result, a promising dice score is achieved on the validation set for the described four-class approach.