Several Dishes Recognition

Learn more about how to preprocess your images in our Image Pre-processing Tutorial page.

Plan & Token Requirements

Feature available in the following LogMeal Plans:

Accessible by the following User Types:

What It Does

Several Dishes Recognition analyzes an image containing multiple foods (like a meal with multiple items in multiple plates, beverages and any type of food container) and splits it into distinct dish regions. Each region is then classified with dish-level predictions, food types and food groups.

When submitting a meal with this feature, its nutritional details will be associated with the user profile performing the request, allowing you to build an image-based food diary.

✕

✕

When to Use It / Outcomes

• You have meal images with more than one dish (e.g., main + sides + drinks).

• You want to separate portions for better tracking, analysis, or dietary reporting.

• You want to submit a new intake for a food diary for later ingredient, quantity and nutrition analysis.

• Output: JSON containing:

  • Bounding boxes and segments/regions for each detected dish region. Use processed_image_size to map bounding boxes to your displayed image size
  • Predictions (dish labels, probability scores)
  • Food type per dish
  • Food groups per dish
  • Personalization‑modulated probability scores using user-specific intakes history (schema unchanged)

Feature-Specific Details

• Uses image segmentation to identify multiple items.

• Each dish is independently classified.

• Results include dish names, confidence scores, associated food type and food groups.

• Can be combined with Ingredients Analysis and Nutritional Analysis.

• Personalization notes:

  • Each submitted/confirmed intake will contribute to the consumption history and thus, modulate the food recognition output.
  • The detected consumption occasion contributes to the history-based food recognition personalization.
  • Minimum history threshold prevents early bias; scores fall back to standard food recognition when not met.
  • No schema changes, only probability values are modulated.

Personalization: Historical Intake Priors

[New] Food recognition predictions will be personalized per end‑user by using historical intake frequency priors. LogMeal subtly adjusts per‑dish probabilities using the user’s confirmed eating history. Priors are kept per meal occasion (e.g., breakfast, lunch, dinner).

How it works

1. User captures a meal and detected dishes update that user’s frequency profile.
2. When confirming/editing the food items in the intake that user’s frequency profile is updated accordingly.
3. On future meals, the profile gently boosts likely dishes; the visual model still drives recognition.

Safeguards

• Cold start: priors activate only after a minimum number of confirmed intakes.
• Light‑touch: probabilities are modulated, never hard‑overridden.
• Auto‑fallback: if there isn’t enough history, the response uses standard visual recognition.
• Privacy: profiles are isolated per APIUser.

Personalization quality depends on sending user confirmations after each meal. See the use case below.

Related Endpoints

Use the following endpoint for multi-dish recognition:

• POST /image/segmentation/complete → 🔴 Performs dish segmentation and returns all dishes detected with predictions and groups.

Check request limits in the endpoint reference. Available to 🔴 APIUser tokens and included in Analyse plans or above.

Related Use Cases

Copy-Paste Recipes

Set profile data once to improve label language and regional dishes:

Send the image to segmentation to obtain regions and top dish candidates per region. Display the top 5 candidate dishes and ask the user for confirmation: