Evaluating Illumination Strategies for Neural-Based Surface Quality Assessment in Cold-Rolled Steel Production

Quality assurance is a fundamental point in the production of cold-rolled metal sheets. Surface defect detection plays a crucial role for components whose aesthetic aspect is important. This work analyses the effect of illumination strategies on detection performances of different neural networks. The illumination of steel sheets from different angles allows highlighting different types of defects. After capturing images under distinct lighting conditions (a grazing light in a dark field setup to enhance fine scratches and a direct diffused illumination in bright field to better reveal larger patches), we performed a multiclass defect classification by training different neural network architectures (EfficientNetB0, EfficientNetB6, EfficientNetV2L and MobileNet). Performances have been quantified using F1-scores for each class. Results showed that scratches are better identified by using dark field models and patches are better detected in bright field ones, regardless of the architecture involved. Neural networks have been compared with anomaly detection methodologies, based on adversarial autoencoders. Anomaly detection provided insightful metrics to gauge the severity of the different defects and allow a comprehensive evaluation of the lamination process.