Steel Surface Defect Classification Using GLCM, SVM and CNN

Hot-rolled steel strip leaves the mill with surface defects such as crazing, inclusions, patches, pitted surface, rolled-in scale and scratches, which decide whether a coil is released, reworked or downgraded. On the public NEU surface defect database (NEU-CLS, Song and Yan 2013: 1,800 grey 200 x 200 camera images, 300 per class) this project classifies each image from its texture - grey-level co-occurrence matrix (GLCM) properties and rotation-invariant local binary patterns (LBP) with a support vector machine (SVM) - and compares it with a small convolutional neural network (CNN). Unlike the related projects (E12 PCB defect detection by template differencing on binarised scan pairs, E11 wafer-map patterns, C1 concrete cracks, E10 traffic signs), PR2 is whole-image TEXTURE classification of grey steel images, with a study of how both approaches degrade under camera noise. On the 360 held-out test images the GLCM + LBP + SVM baseline reaches 94.72 % and the small CNN 94.44 % - a statistical tie (McNemar p = 1.0); both collapse under mild Gaussian noise, and an SVM trained on noisy copies keeps about 90 % and was deployed. A FastAPI web application with five interchangeable databases classifies images, keeps mill lines, coils, inspections and the inspector's review; GNU Octave and ImageJ reproduce the features.

Category: Production Engineering

Production Engineering Project PR02 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.