SteelNet

The Problem

Industrial steel defect detection requires reliable, automated quality control. Traditional manual inspection is slow and error-prone. The challenge: can modern deep learning outperform classical computer vision approaches for this task?

My Approach

Designed and ran a rigorous comparative experiment between a fine-tuned ResNet18 CNN and a traditional HOG+SVM pipeline on the NEU-DET steel surface defect dataset. Built a complete ML pipeline including data augmentation, cross-validation, hyperparameter tuning, and comprehensive evaluation metrics. Focused on reproducible research methodology, not just model accuracy.

Key Results

• HOG+SVM achieved 92.44% accuracy and 92.31% F1 score
• ResNet18 CNN achieved competitive performance with room for further tuning
• Published comparative analysis with confusion matrices and per-class metrics
• Top final project score in CSC 537 Deep Learning course

What I Learned

Classical methods can still compete with deep learning when the dataset is well-structured and features are domain-appropriate. The research process — designing experiments, controlling variables, drawing defensible conclusions — matters as much as the final accuracy number.