food-image-classification

🍽️ Food Image Classification

This project applies Convolutional Neural Networks (CNNs) and transfer learning to classify food and vegetable images.
It was designed as a hands-on computer vision project to practice data collection, preprocessing, model training, and evaluation.

---

🎯 Objectives

• Build a food image classifier using CNNs and transfer learning
• Practice end-to-end CV workflow: data crawling, preprocessing, augmentation, training, evaluation
• Understand the challenges of multi-class classification with real-world image data

---

📂 Dataset

• Source: Collected via Google Image crawling (for educational and non-commercial use)
• Categories: Fruits and vegetables (~50 classes)
• Availability: Raw images are excluded from the repository to avoid copyright issues

⚠️ All images are collected for educational purposes only.
If you want to build a similar dataset, use the crawling notebook in notebooks/01_food-image-crawling.ipynb.

---

🧠 Methods

• Modeling:
  • Baseline CNN
  • Transfer learning with MobileNetV2 (ImageNet pretrained)
• Training setup:
  • Optimizer: Adam
  • Learning rate scheduler, EarlyStopping, ModelCheckpoint
  • Epochs: 30
• Data Augmentation: Resizing, normalization, random flips/rotations

---

📊 Results

Training Curves

Confusion Matrix

Evaluation Report

• Validation Accuracy: ~61% (Top-1)
• Top-3 Accuracy: ~80%+
• Best classes: banana, kiwi, pineapple (F1 > 0.8)
• Challenging classes: apple, capsicum, corn

---

🧱 Project Structure

food-image-classification/
├── figures/                          # Evaluation plots & confusion matrix
├── images/                           # Crawled dataset (excluded in .gitignore)
├── notebooks/                        # Experiment notebooks
│   ├── 01_food-image-crawling.ipynb        # Image crawler with Selenium
│   └── 02_food-image-classification.ipynb  # Model training & evaluation
├── best_model.h5                     # Trained model weights (excluded in .gitignore)
├── .gitignore
└── README.md

🚫 Limitations

• Dataset imbalance — some classes contain far fewer images than others
• Visual similarity between certain categories (e.g., cabbage vs lettuce, capsicum vs chili pepper) reduces classification accuracy
• Crawled images include noise (low quality, mislabeled, duplicates) which affects model generalization
• Achieved validation accuracy (~61%) is modest and not yet sufficient for production-level use

---

🔭 Next Steps

• Experiment with more advanced architectures (EfficientNet, Vision Transformers)
• Apply stronger augmentation strategies (CutMix, Mixup, AutoAugment)
• Investigate explainability tools (Grad-CAM, SHAP) to understand model decisions
• Deploy a lightweight demo app with Streamlit or Flask for interactive testing

---

👤 Author

Maintained by hojjang98
📅 Last updated: September 2025