Robust Deep Neural Network for Learning in Noisy Multi-Label Food Images.

Area of Science:

• Computer Science
• Artificial Intelligence
• Nutrition Science

Background:

• Deep learning models require large, clean datasets for accurate food recognition, crucial for monitoring balanced diets and preventing eating disorders.
• Manual data cleaning for food datasets is a significant bottleneck in developing effective dietary monitoring systems.

Purpose of the Study:

• To develop a deep learning method capable of learning from noisy multi-label food data, reducing reliance on extensive data cleaning.
• To extend the Attentive Feature MixUp technique to robustly handle imperfections in food image datasets.

Main Methods:

• The Attentive Feature MixUp method was extended with a hypothesis-driven approach to label generation during image mixing.
• Investigated the incorporation of Laplace approximation as a post-hoc method to further address data perturbation.
• Evaluated the proposed method on two distinct food datasets.

Main Results:

• The extended Attentive Feature MixUp method demonstrated significant performance improvements on food datasets.
• Achieved notable gains in Jaccard index and F1 score, validating the effectiveness of the label generation hypothesis.
• The method successfully reduced the memorization of noisy multi-labels, leading to enhanced performance.

Conclusions:

• The proposed method enables deep learning models to learn effectively from noisy multi-label food data, a critical step towards reliable dietary monitoring.
• This work offers a practical solution to the challenge of data cleaning in food image analysis for health applications.
• The findings support the use of advanced MixUp strategies for building more resilient AI systems in nutrition and health.