Manufacturing Defect Classifier

A binary image classifier that flags defective vs. acceptable manufactured parts, built with transfer learning on a pretrained ResNet18, and served through a Flask web app for interactive predictions.

Overview

This project fine-tunes a pretrained ResNet18 on real casting images to classify each part as defective (def_front) or acceptable (ok_front), then wraps the trained model in a simple web interface for live predictions on new images.

Results

• Test accuracy: 99.72%
• Trained on 6,633 labeled casting images (3,758 defective, 2875 acceptable), evaluated on a held-out test set of 715 images
• Test loss and accuracy plateaued by epoch 3-4, indicating the model converged quickly on this dataset

Epoch	Train Loss	Train Acc	Test Loss	Test Acc
0	0.4404	0.8274	0.2204	0.9077
1	0.3014	0.8874	0.0779	0.9804
2	0.2171	0.9144	0.0185	0.9944
3	0.1026	0.9638	0.0192	0.9958
4	0.1053	0.9619	0.0111	0.9972

Precision, Recall & Confusion Matrix

Recall on the def_front class matters more here than overall accuracy, since missing a real defect is more costly than a false alarm.

	precision	recall	f1-score	support
def_front	1.00	1.00	1.00	453
ok_front	0.99	1.00	1.00	262
accuracy			1.00	715
macro avg	1.00	1.00	1.00	715
weighted avg	1.00	1.00	1.00	715

Dataset

Real-life Industrial Dataset of Casting Product — submersible pump impeller images, labeled as defective or acceptable, pre-split into train/test sets. Not included in this repo (see Setup below to download it yourself).

Approach

Rather than training a CNN from scratch, this project uses transfer learning: starting from a ResNet18 pretrained on ImageNet, then replacing its final classification layer and fine-tuning the entire network on the casting dataset. Since pretrained models already encode general visual features (edges, textures, shapes) that transfer well to new domains, this means far less data and training time is needed than training from scratch.

Pipeline:

1. Load and preprocess casting images (resize, normalize to ImageNet statistics)
2. Replace ResNet18's 1000-class output layer with a 2-class layer (defective / ok)
3. Fine-tune all layers using SGD with a step-decay learning rate schedule
4. Track train/test loss and accuracy per epoch; save the best-performing checkpoint
5. Serve the trained model through a Flask API with a browser-based upload interface

Tech Stack

• ML: PyTorch, torchvision (ResNet18, transfer learning)
• Backend: Flask
• Frontend: HTML/CSS (server-rendered via Jinja2 templates)

Project Structure

├── main.py                  # Training orchestration: builds model, trains, saves weights
├── model.py                 # ResNet18 architecture definition
├── train.py                 # Training loop (train/test phases, checkpointing)
├── data.py                  # Dataset loading and transforms
├── evaluate.py              # Precision/recall/confusion matrix evaluation
├── run_evaluation.py        # Loads saved weights and re-runs evaluation without retraining
├── config.py                # Device configuration (CPU/GPU)
├── visualize.py             # Visualizes sample predictions after training
├── app.py                   # Flask app for interactive inference
├── templates/
│   └── index.html           # Upload form + prediction display
├── model_weights.pth        # Trained model weights
└── requirements.txt

Setup

1. Clone and install dependencies:

git clone https://github.com/Hosam-Issa/Manufacturing-Defect-Classifier
cd Manufacturing-Defect-Classifier
python -m venv venv
venv\Scripts\activate          # Windows
pip install -r requirements.txt

Note on PyTorch/CUDA: if you have an NVIDIA GPU and want GPU-accelerated training, install PyTorch separately first using the correct command for your CUDA version from pytorch.org/get-started/locally, before running the command above. The CPU-only build works fine for running inference (the Flask app) even without a GPU.

2. Download the dataset (only needed to retrain the model since model_weights.pth is already included):

• Download from the Kaggle link above
• Extract into an archive/ folder in the project root, matching the path structure expected in data.py 3. Run the web app:

python app.py

Visit http://127.0.0.1:5000, upload a casting part image, and get an instant defect prediction.

4. (Optional) Retrain the model:

python main.py

5. (Optional) Re-run evaluation without retraining:

python run_evaluation.py

Future Improvements

• Multi-class classification for different defect types, rather than binary defective/ok
• Add prediction confidence scores alongside the classification
• Expand to a general-purpose defect detection tool across other manufacturing datasets