Automatic Detection and Quantification of Hot-Rolled Steel Surface Defects Using Deep Learning

Timely defect detection plays a positive role in preventing the evaluation of steel derivative damage. As a state-of-the-art image recognition technology, pixel-level segmentation technology can obtain the pixel distribution and generate the shape of the objects accurately, which provides a potential detection method for steel surface defects. Therefore, this paper employed the well-known pixel-level segmentation CNN (DeepLab_v3+) to detect three defect categories (inclusions, patches and scratches) on the steel surface, and the ‘ regionprops ’ function was used to quantify the defect features (length, average width, maximum width, area and ratio). The results show that ResNet50, as the backbone network of DeepLab_v3+, has the highest detection precision for steel surface defects, and its accuracy is more competitive than that of other algorithms (FCN, SegNet, U-Net and PAG-Net). The proposed quantitative method also achieved encouraging results (the average relative error (ARE) of the evaluation indicator, 10%, 18%, 17%, 23% and 23%, respectively), and the precision was higher than that of the other methods. This demonstrated that the proposed method can greatly benefit steel surface defect detection and evaluation of defect levels.