mobilenet-food-binary

MobileNetV3 Food Classifier (3-class)

A lightweight, production-ready image classifier for food content moderation. Classifies images into healthy food, unhealthy food, or not-food (non-meal content).

Model Details

• Architecture: MobileNetV3Small (frozen backbone) + trainable classification head
• Input: 224x224 RGB images
• Output: 3 classes (healthy, unhealthy, not_food)
• Framework: TensorFlow/Keras
• Best Validation Accuracy: 92.92%

Performance

Per-Class Performance

Training Configuration

Stage 1: Frozen Backbone (20 epochs)

• ImageNet-pretrained MobileNetV3Small with frozen weights
• Classifier head: GlobalAveragePooling2D -> Dropout(0.2) -> Dense(3, softmax)
• Optimizer: Adam (lr=1e-3)
• Loss: Categorical Crossentropy
• Best val accuracy: 91.56%

Stage 2: Fine-tuning (3 epochs)

• Unfroze backbone, kept 52 BatchNorm layers frozen
• Optimizer: Adam (lr=1e-5)
• Fine-tuned for food-specific feature learning
• Final val accuracy: 92.92%

Data Augmentation

• RandomFlip (horizontal)
• RandomRotation (0.05)
• RandomZoom (0.1)

Dataset

• Total: 9,714 validation images from Food-101
• Healthy (3,716): Fruits, grains, salads, seafood, smoothies, soups, vegetables
• Unhealthy (2,998): Burgers, candy, desserts, fried food, pizza, snacks, sugary drinks
• Not-Food (3,000): General objects, street numbers, screenshots

Model Variants

Uncertainty Analysis

Confidence threshold = 0.70:

• Certain predictions: 90.8%
• Routed to review: 9.2%
• Accuracy on certain: 94.33%

License

Apache 2.0