shashikantkaushik/Surface-defects-classification-of-the-hot-rolled-steel-strip

Surface Defects Classification of the Hot-Rolled Steel Strip

A ResNet50-based deep learning model for automated visual inspection of hot-rolled steel surfaces, built using the Aargus DIY visual inspection tool. The model classifies steel surface defects with 97.50% accuracy.

Note: This release covers 4 defect classes (crazing, inclusion, patches, scratches). A 6-class version (adding pitted_surface and rolled-in_scale) is in progress — see Roadmap.

Overview

This model performs automated visual inspection of hot-rolled steel surfaces to detect and classify manufacturing defects, replacing slow and inconsistent manual inspection with a fast, consistent AI-based system. It was built using the Aargus DIY visual inspection tool — covering data ingestion, augmentation, transfer learning, and statistical validation.

Defect Classes: crazing, inclusion, patches, scratches

Methodology

1. 1.Data Ingestion — Stratified 80/20 train-validation split
2. 2.Preprocessing & Augmentation — rotation, horizontal/vertical flip, zoom, and brightness variation applied to improve generalization
3. 3.Model Architecture — ResNet50 backbone (ImageNet pretrained), trained in two phases:
4. 4.Phase 1: frozen base warm-up (13 epochs)
5. 5.Phase 2: selective fine-tuning of the last 60 layers (6 epochs)
6. 6.Class Balancing — computed class weights to address any dataset imbalance
7. 7.Validation — classification metrics (Precision, Recall, F1), confusion matrix analysis, ROC-AUC curves, precision-recall curves, and bootstrap confidence intervals for statistical robustness

Performance

Overall Metrics

Per-Class Results

Total misclassified: 6 out of 240 validation images

Confidence Analysis

The large gap between correct and incorrect prediction confidence indicates the model is well-calibrated — it tends to be less confident when it makes a mistake, rather than being confidently wrong.

Visual Results

Confusion Matrix

[image]

Training Curves (Accuracy & Loss)

[image]

ROC Curve — Per Class

[image]

Precision-Recall Curve — Per Class

[image]

Prediction Confidence Distribution

[image]

Calibration Curve

[image]

Bootstrap Accuracy Distribution (95% CI)

[image]

Misclassified Examples

[image]

Usage

Installation

pip install tensorflow pillow numpy

Load the model and predict

import tensorflow as tf
from tensorflow.keras.applications.resnet50 import preprocess_input
import numpy as np
from PIL import Image

# Load model
model = tf.keras.models.load_model("final_model.h5")
class_names = ["crazing", "inclusion", "patches", "scratches"]

# Load and preprocess an image
img = Image.open("your_steel_surface_image.jpg").resize((224, 224))
arr = preprocess_input(np.array(img))
arr = np.expand_dims(arr, axis=0)

# Predict
preds = model.predict(arr)[0]
predicted_class = class_names[np.argmax(preds)]
confidence = np.max(preds)

print(f"Predicted defect: {predicted_class} ({confidence*100:.2f}% confidence)")

Use Case

Industrial quality control automation for steel manufacturing — reduces manual inspection time and improves defect detection consistency across production lines.

Roadmap

• —[ ] Retrain on full 6-class dataset (add pitted_surface, rolled-in_scale)
• —[ ] Publish updated metrics and confusion matrix for 6-class model
• —[ ] Deploy as REST API endpoint

Links

Built With

• —TensorFlow / Keras
• —ResNet50 (ImageNet pretrained)
• —scikit-learn (evaluation metrics)
• —Google Colab (training environment)

License

This project is licensed under the Apache 2.0 License.

Author

Aargus — DIY Visual Inspection AI Platform