Reproduce “Detecting Korean Food Using Image using Hierarchical Model” on OpenResearch

About this paper

Hoang Khanh Lam, Kahandakanaththage Maduni Pramuditha Perera

Researchers at Dong-A University developed a hierarchical model for identifying Korean food dishes from images, primarily assisting individuals with dietary restrictions or language barriers. This system, employing multi-stage hierarchical transfer learning with YOLOv8, achieved an average classification accuracy of 88.50%, improving upon a flat classification baseline by 3.18 percentage points.

Key insights

• A hierarchical data structure and classification approach can effectively address issues like class imbalance and subtle visual distinctions, leading to improved food recognition accuracy.
• Multi-stage transfer learning, by first classifying broad food types and then refining to specific dishes, manages classification complexity and enhances overall model performance.
• The inherent interpretability of hierarchical models provides a clearer understanding of relationships between different food categories, aiding in system design and user navigation.

Summary by alphaXiv.