Image-Based Dish Recognition Through Thematic Variant Matching

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing diet logging systems require significant user interaction and have difficulty accurately recognizing diverse dishes due to overlapping class definitions and variations in food items, leading to frustration and inaccurate classification.

Innovation Solution

An image-based dish recognition method using machine learning and deep learning processes that analyze a dish image, cluster recipes based on similarity, and calculate association scores to identify centroid and variant dishes, enhancing user-friendliness and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If classes are defined at fine granularity to achieve high specificity, then dish recognition specificity is improved, but class specification becomes difficult and training data collection becomes complex

Engineering Contradiction:
Improvedish recognition specificityVSAvoidclass specification complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the dish recognition task into two stages: first performing image recognition at a coarse thematic level to identify broad dish categories, then performing natural language processing on recipe descriptions to distinguish specific variants. This segmentation allows the system to achieve high specificity without the complexity of defining fine-grained classes from scratch, as the thematic level provides a simplified classification structure.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces recipe descriptions and natural language processing as an intermediary between image recognition and final dish identification. The recipe text acts as a mediator that bridges the gap between visual appearance (which may be similar across variants) and precise dish identification, allowing the system to achieve high specificity without requiring complex visual differentiation for each dish variant.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Device complexity

If classes are defined distantly apart to avoid overlapping, then class definition simplicity is improved, but dish recognition sensitivity decreases

Engineering Contradiction:
Improveclass definition simplicityVSAvoiddish recognition sensitivity
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the recognition process into thematic classification (coarse level) and variant identification (fine level). The thematic level uses simple, non-overlapping categories that are easy to define, while the variant identification stage uses recipe text analysis to achieve high sensitivity. This segmentation allows the system to maintain simple class definitions while still achieving high recognition accuracy through the two-stage process.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent adds a textual dimension to the recognition process by analyzing recipe descriptions. This textual dimension provides an additional feature space that helps distinguish between dishes with similar appearances. By incorporating this additional dimension, the system can achieve high sensitivity without requiring complex visual differentiation, as the text provides discriminative information that complements the visual features.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If machine learning processes are used to extract patterns from training data, then dish recognition accuracy is improved, but the system becomes vulnerable to overlapping class definitions and data collection challenges

Engineering Contradiction:
Improvedish recognition accuracyVSAvoiddata collection complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent uses recipe descriptions as an intermediary that simplifies the data collection and training process. Instead of requiring extensive visual data for each fine-grained class, the system can leverage existing recipe text data to train the natural language processing component. This intermediary approach reduces the complexity of data collection while maintaining high recognition accuracy, as the recipe text provides rich semantic information that complements visual features.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent makes the recipe description serve multiple functions: it provides dish identification, offers cooking instructions, and supplies semantic information for the natural language processing stage. This multi-functionality reduces the need for separate data collection efforts for different purposes, thereby reducing overall data collection complexity while maintaining high accuracy through the combined visual and textual analysis.

Inventive Principle:
Principle #6Universality (Multi-functionality)

4Measurement precision

If users manually input and record consumed dishes, then recording accuracy can be controlled, but user interaction requirement increases and adherence decreases

Engineering Contradiction:
Improverecording accuracyVSAvoiduser interaction requirement
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The patent implements self-service by using image recognition to automatically identify dishes without requiring manual input from users. The system captures images of consumed dishes and automatically processes them through the two-stage recognition process (image analysis followed by recipe text analysis) to identify the dish. This eliminates the need for users to manually record what they ate, significantly reducing user interaction requirements while maintaining high accuracy through automated pattern recognition and natural language processing.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3948650B1An apparatus and method for performing image-based dish recognition
Publication Date: 2025.07.23 VERSUNI HLDG BV
  • EP3948650B1 patent drawingFigure 1
  • EP3948650B1 patent drawingFigure 2
  • EP3948650B1 patent drawingFigure 3A~3B

AI summary

There is provided a computer-implemented method for performing image- based dish recognition. The method comprises: acquiring (202) a first image depicting a dish to be recognized; analyzing (204) the first image using a predictive model to determine a first candidate theme; acquiring (206) a reference set of dish identifiers; calculating (208), for each of the dish identifiers in the reference set, an association score indicative of a degree of similarity between the dish represented by the respective dish identifier and the first candidate theme; selecting (210) one or more dish identifiers in the reference set having the highest association scores as one or more variant dish identifiers of the first candidate theme; and outputting (212) the centroid dish identifier of the first candidate theme and the one or more variant dish identifiers of the first candidate theme.