Adaptation of Models for Food Intake Sound Recognition Using Maximum a Posteriori Estimation Algorithm

Obesity and overweight are big healthcare challenges in the world's population. Automatic food intake recognition algorithms based on analysis of food intake sounds offer the potential of being a useful tool for simplifying data logging of consumed food. High inter-individual differences of the users' food intake sounds decrease the classification accuracy achieved with a user-unspecific algorithm. To overcome this problem, the Maximum a Posteriori (MAP) estimation is implemented and tested on one user consuming eight types of food. The dependency of the classification enhancement from the size of the adaptation set is investigated. Overall recognition accuracy can be increased from 48 % to around 79 % using records of 10 intake cycles for every food type of one subject. An increase by 7.5 % can be shown for a second subject. This shows the usability of the MAP adaptation algorithm at food intake sound classification tasks. The algorithm provides a suitable way for adapting models to a user, thereby, enhancing the performance of food intake classification.