Food Recognition System: A New Approach Based on Wavelet-LSTM

Abstract

An automated system for analyzing daily dietary intake is essential for human well-being and healthcare. This work presents a novel wearable necklace embedded with a piezoelectric sensor and a microcontroller to monitor food ingestion of users. To effectively represent the food ingestion patterns, the sensor signal is dynamically segmented using a bidirectional search technique. Each segmented food intake pattern consists of a chewing sequence and a swallow peak. We exploit wavelet transform to decompose the complex food ingestion patterns, collected by the sensor, into frequency sub-bands at discrete scales. The frequency sub-bands are used as sequences to train long short-term memory (LSTM) for the recognition of 5 food categories. Our proposed recognition model based on wavelet-LSTM recognizes 5 food classes with an accuracy of 98.1%