NutriVision: The AI Dietitian in Your Pocket

1. The Spatial Anchor

By simply placing a 1-rupee coin (21.93 mm in diameter) next to a plate, the system uses the coin as a reference point. This allows it to calculate the exact volume and weight of the food, bridging the gap between a simple picture and a clinical data point.

2. From Image to Volume

To turn a 2D image into 3D nutritional data, the team applied a 0.8 compensation factor. This factor accounts for the empty space within the food’s rectangular bounding boxes. The calculated volume is then translated into grams.

3. The Intelligent Analysis Engine

The processed data is filtered through an NLP-driven chatbot. If a user’s clinical profile suggests they are at risk (e.g., for diabetes), the system doesn't just record the meal. It triggers a real-time, life-saving warning regarding high sugar or carbohydrate content.

Solving "Under-Reporting"

Current health apps suffer from massive "under-reporting"—humans are notoriously bad at estimating how much they actually eat. NutriVision removes this cognitive burden entirely.

AI Model Performance

The system's engine is a Faster R-CNN with a ResNet-50 backbone, which was found to be significantly more reliable than other models like YOLO.
In rigorous testing:

• Classification Accuracy: 92%
• Precision: 82%
• High Confidence Examples: A slice of pizza (99.32% confidence) or a piece of cake (99.5% confidence). The AI doesn't just see "food"—it calculates its mass.

Technical Constraints

While a leap forward, the architecture has some current limitations:

• It requires the presence of the specific 21.93 mm reference coin.
• Model confidence can waver when food items are stacked or overlapped too densely.
• The development dataset was limited to 500 images (plus 200 for validation).

The Next Steps

To move from the lab to global kitchens, the system needs to:

• Expand beyond its current 10 food classes.
• Learn to account for "post-prandial" leftovers (the food we leave behind on the plate).