Sequential transfer learning based on hierarchical clustering for improved performance in deep learning based food...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate food segmentation from optical images is crucial for applications like nutrition intake monitoring.
• Deep learning models face challenges in achieving human-level accuracy due to the vast diversity of food choices.
• Existing methods struggle with variations in food appearance and presentation.

Purpose of the Study:

• To propose a novel Sequential Transfer Learning method combined with Hierarchical Clustering for enhanced food image segmentation.
• To improve the accuracy and robustness of automated food identification systems.
• To address the limitations of current deep learning solutions in handling diverse food datasets.

Main Methods:

• A novel Sequential Transfer Learning approach was developed, integrating Hierarchical Clustering.
• The method simulates a step-by-step problem-solving framework by clustering similar food types.
• The model was trained and evaluated on optical food images, including a specific application for Danish school children's meals.

Main Results:

• The proposed method achieved up to a 6% gain in accuracy compared to traditional deep learning training.
• The developed model demonstrated improved performance on challenging, previously unseen food image segmentation tasks.
• The approach proved effective in segmenting foods for dietary intake monitoring in a real-world application.

Conclusions:

• Sequential Transfer Learning with Hierarchical Clustering offers a significant improvement for food image segmentation.
• This method enhances the accuracy and robustness of automated food identification, particularly for nutrition monitoring.
• The approach shows promise for practical applications, such as analyzing dietary intake in specific populations.