Abstract

Accurate food intake monitoring is essential for maintaining a healthy diet and preventing nutrition-related diseases. Traditional food classification models struggle with the diverse range of foods across cultures and the continuous introduction of new food types due to relying on fixed-sized datasets. Moreover, studies show that people consume only a small range of foods across the existing ones. These limitations necessitate the model to adapt itself as new classes appear. Additionally, the model needs to pay more attention to certain food classes.
While existing class-incremental models have low accuracy for the new classes and lack personalization, this work introduces a personalized, class-incremental food classification model designed to address these challenges and enhance food intake monitoring systems. Our approach dynamically adapts to new real-world food classes, maintaining accuracy for both new and existing classes through personalization. The model prioritizes foods based on individuals eating habits by considering meal frequencies, times, and locations.
We employ a modified dynamic support network (DSN), the personalized dynamic support network (PDSN), to handle new food classes and propose a comprehensive framework integrating this model into a food intake monitoring system. The system analyzes meal images, estimates food weight with a smart scale, calculates macro-nutrient content, and creates a dietary profile via a mobile app. Experimental evaluations on the personalized datasets FOOD101-Personal and VIPER-FoodNet-Personal (VFN-Personal) demonstrate the model's effectiveness in improving classification accuracy, addressing the limitations of conventional and class-incremental food classification models.