Surface defects classification of the hot-rolled steel strip

About Model

This model performs automated visual inspection of hot-rolled steel surfaces to detect and classify manufacturing defects. It was built using the Aargus DIY visual inspection tool. Methodology: 1. Data Ingestion: Surface Defect Dataset, stratified 80/20 train-validation split 2. Preprocessing & Augmentation: rotation, horizontal/vertical flip, zoom, and brightness variation applied to improve generalization 3. Model Architecture: ResNet50 backbone (ImageNet pretrained), trained in two phases — frozen base warm-up followed by selective fine-tuning of upper layers 4. Class Balancing: computed class weights to address any dataset imbalance 5. Validation: classification metrics (Precision, Recall, F1), confusion matrix analysis, ROC-AUC curves, and bootstrap confidence intervals for statistical robustness Defect Classes: crazing, inclusion, patches, scratches Performance: - Accuracy: 97.50% - Matthews Correlation Coefficient: 0.9674 - 95% Confidence Interval (bootstrap): [95.42%, 99.17%] Use Case: Industrial quality control automation for steel manufacturing — reduces manual inspection time and improves defect detection consistency.