🍔 FoodVision: Deep Learning Food Classification

A computer vision system that classifies 75,000+ food images across 101 categories using Convolutional Neural Networks and Vision Transformers.

🎯 Key Features

• 77% top-1 accuracy on 101 food categories
• Processes 75,000+ training images
• Transfer learning with EfficientNetB0
• Systematic data augmentation (rotation, scaling, color jittering)
• Comprehensive evaluation pipeline with per-class metrics

🛠️ Tech Stack

• Framework: PyTorch, TensorFlow
• Architecture: EfficientNetB0 (Vision Transformer)
• Libraries: torchvision, scikit-learn, pandas, NumPy, matplotlib
• Techniques: Transfer learning, data augmentation, regularization (dropout, weight decay)

📊 Results

Metric	Score
Top-1 Accuracy	77%
Training Images	75,000+
Categories	101
Validation Images	25,000

Performance Highlights

• Reduced overfitting by 23% through regularization
• 15% improvement over baseline CNN
• Per-class precision/recall analysis with confusion matrix

🔍 Challenges & Solutions

• Challenge: High visual diversity across categories
• Solution: Systematic augmentation + transfer learning
• Challenge: Similar food types (e.g., pasta varieties)
• Solution: Fine-grained feature extraction focus

📝 Key Learnings

• Transfer learning dramatically improves performance on limited data
• Data augmentation must preserve category-defining features
• Per-class analysis reveals model strengths/weaknesses better than overall accuracy

📧 Contact

Romeo Nickel - LinkedIn - rjnickel@usc.edu