Multi-Scale Multi-View Deep Feature Aggregation for Food Recognition.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Food recognition is crucial for health applications but current methods using Convolutional Neural Networks (CNNs) struggle with unique food image characteristics.
• Food images lack distinct spatial arrangements and semantic patterns, hindering traditional object recognition approaches.
• Existing methods often overlook the specific visual properties of food, limiting recognition accuracy.

Purpose of the Study:

• To propose a novel Multi-Scale Multi-View Feature Aggregation (MSMVFA) scheme for enhanced food recognition.
• To develop a unified feature representation by aggregating semantic, attribute, and visual features.
• To improve the robustness and discriminative power of food recognition systems.

Main Methods:

• Utilized ingredient knowledge with ingredient-supervised CNNs for mid-level attribute features.
• Extracted high-level semantic and deep visual features using class-supervised CNNs.
• Implemented multi-scale CNN activation fusion and multi-view feature aggregation for a comprehensive representation.

Main Results:

• Achieved state-of-the-art Top-1 recognition accuracy on three large-scale food benchmark datasets.
• Demonstrated superior performance compared to existing food recognition methods.
• The MSMVFA scheme proved robust and effective across diverse food image datasets.

Conclusions:

• The proposed MSMVFA scheme offers a more robust and comprehensive approach to food recognition.
• This method effectively captures food image semantics by integrating multi-granularity features.
• The findings are expected to advance the field of food recognition in computer vision.