Ingredient Prediction via Context Learning Network With Class-Adaptive Asymmetric Loss

Area of Science:

• Computer Vision
• Machine Learning
• Food Science

Background:

• Ingredient prediction from food images is crucial for applications like nutrition tracking.
• Current methods often overlook ingredient-specific features and struggle with imbalanced data distributions.
• Existing approaches primarily focus on joint learning, neglecting the independent characteristics of ingredients.

Purpose of the Study:

• To propose a novel framework, Class-Adaptive Context Learning Network (CACLNet), for enhanced ingredient prediction.
• To improve feature extraction by considering comprehensive and detailed ingredient characteristics.
• To address the challenge of extreme class imbalance in ingredient datasets.

Main Methods:

• Introduced Ingredient Context Learning (ICL) for self-supervised feature extraction, reducing background noise and enhancing ingredient region connections.
• Developed Class-Adaptive Asymmetric Loss (CAAL) to adaptively focus on different ingredient classes and manage positive-negative sample imbalance.
• Evaluated CACLNet on Vireo Food-172 and UEC Food-100 benchmark datasets.

Main Results:

• CACLNet achieved state-of-the-art performance on both Vireo Food-172 and UEC Food-100 datasets.
• Ingredient Context Learning effectively reduced background interference and strengthened ingredient feature representation.
• Class-Adaptive Asymmetric Loss successfully handled class imbalance, improving prediction accuracy for rare ingredients.

Conclusions:

• The proposed CACLNet framework significantly advances ingredient prediction accuracy.
• ICL and CAAL are effective components for improving feature extraction and addressing data imbalance in food ingredient recognition.
• The method demonstrates strong potential for real-world applications in food analysis and management.