RatioLogo
Back

The AI and the Invisible Diet

What if the secret to modern longevity—the Mediterranean diet—is effectively invisible to the very artificial intelligence meant to track our health? For years, computer vision systems have been trained on generic burgers and pizzas, leaving a significant clinical blind spot for regional, nutrient-dense cuisines like those found in Catalonia.

Bridging the Digital Divide with FoodCAT

Researchers at the University of Barcelona are addressing this gap with FoodCAT, a novel dataset designed to teach AI the nuances of the Catalan kitchen.

The Dataset

This is more than a culinary catalog; it is an essential tool for "lifelogging." The dataset allows clinicians to:

  • Automatically monitor dietary habits.
  • Identify unhealthy patterns with surgical precision.

Overcoming Technical Hurdles

The team faced a significant challenge: real-world food photos are often messy and low-resolution.

The Solution: Super-Resolution

To combat poor image quality, researchers employed Super-Resolution (SR) technology.

  • They used a Sparse Coding-based Network to upscale images.
  • Images were enhanced to a crisp 256x256 pixels for processing.

Striking Results in Recognition

When these enhanced images were processed, the AI model achieved impressive accuracy.

Model Performance

The model, processed through GoogleNet, delivered strong results:

  • Top-5 accuracy: 97.07% for broad food categories.
  • Top-1 accuracy: 68.07% for specific dish recognition.
  • Accuracy was measured across a combined 216 different food classes.

Key Insights on Training "Digital Eyes"

The study revealed that how we train AI matters as much as what it sees.

Fine-Tuning vs. Training Time

A critical finding was the impact of deep network adjustment:

  • Fine-tuning every layer of the neural network (not just the final one) boosted accuracy by 10.93%.
  • This deeper "thinking" process doubled the training time to 24 hours.

The Power of Data Volume

Performance was heavily influenced by the size of the training category:

  • "Desserts and sweets" (the largest category with 11,933 images) performed significantly better.
  • Smaller categories like "Mushrooms" (with only 438 images) showed lower recognition rates.
  • This highlights that volume is king for model accuracy.

Current Limitations and Future Potential

While highly effective, the system has room for growth and faces specific challenges.

Identified Challenges

Researchers noted several areas for improvement:

  • Data Disparity: High variance in image counts between categories can skew performance metrics.
  • Localized Expertise: The model's knowledge is currently specific to Catalonia.
  • Training Limits: Even higher accuracy may be possible if models train beyond 1,000,000 iterations.

Conclusion: A New Data Point for Health

For now, the FoodCAT project proves that AI can learn the visual language of regional health. It turns a simple smartphone photo into a sophisticated data point, paving the way for better tools in cardiovascular disease prevention and dietary monitoring.

This summary is based on the research paper: "Can a CNN Recognize Catalan Diet?" by Pedro Herruzo, Marc Bolaños, and Petia Radeva (University of Barcelona; Computer Vision Center), published July 2016.