This repository contains deep learning models for food classification, covering binary, 10-class, and 101-class classification tasks. The models are trained using custom CNN architectures and transfer learning with EfficientNet and ResNet.
🔗 Dataset: FoodVision Dataset Link
🔗 Live Training Logs & Results: Tracked with Weights & Biases (WandB)
The goal of this project is to develop an efficient food classification system capable of distinguishing between different food categories. The following tasks were performed:
✅ Binary Classification: Pizza 🍕 vs. Steak 🥩
✅ 10-Class Classification: Classifying images into 10 different food categories
✅ 101-Class Classification: Classifying images from the full Food101 dataset
✅ Transfer Learning: Using EfficientNet & ResNet with Feature Extraction & Fine-Tuning
✅ Training on 10% vs. 100% of Data: Evaluating model performance on limited vs. full dataset
✅ Data Augmentation: Improving generalization using transformations like flipping, rotation, and zoom
- EfficientNetB4 (Fine-Tuned) performed best for 101-class classification on 100% of the data.
- Feature extraction works well on small datasets, but fine-tuning improves performance on large datasets.
- Data augmentation helped improve model generalization, reducing overfitting.
🔹 Experiment with Swin Transformers for food classification.
🔹 Optimize training using learning rate schedules.
🔹 Deploy model as a web application for real-time classification.
👤 Muhammad Hamza Azhar
If you found this project useful, please consider starring 🌟 the repository and sharing it!
Happy Coding! 🚀