Detection of Defective Steel Surface with Image Segmentation

The detection of defects is an important part of steel manufacturing. Conventional ways of detecting defects in steel include ultrasonic testing, magnetic powder testing, osmosis testing, eddy current testing and X-ray testing. In this study, we used deep learning and computer vision to identify surface defects on steel. The study used U-Net, Mask R-CNN and Xception-V1which are deep neural network models used to detect specific parts of an image or image segmentation. The Xception- V1model performed well with complex defects whereas Mask-RCNN gave better results in rare defects. U-Net performed well with every type of defect and was a versatile model. The accuracy of U-Net model came out to be 80.21%. Images with no defects resulted in higher percentage of false predictions. It can be concluded that the existing models, while are capable of identifying different defects, they are not very good at predicting if there is no defect in this particular case.