Modelo de visão computacional para identificação de alimentos em refeições a partir da segmentação obtida por múltiplas redes completamente convolucionais

Abstract

The strategy of counting carbohydrates in consumed foods is recommended by scientific societies as a way to improve the quality of life for diabetes patients. Monitoring food intake can be facilitated by using a mobile application that automatically recognizes the foods in a meal. Automatic recognition of food images is considered a challenging task for computer vision due to the similarity between foods. This challenge increases when the goal is to classify foods from a specific region and with a dataset containing only foods from that region, and therefore, small compared to public datasets from other countries. For this task, this work presents a model that uses a set of Fully Convolutional Networks (FCNs) to generate segmentations of foods in a meal. These segmentations are processed by an algorithm that classifies the foods using digital image processing techniques. The model has low training costs and is scalable, meaning it can be trained to recognize a new food without the need to retrain the entire model. In the tests, foods consumed in Brazil were used, achieving an accuracy of 98.2% and a recall of 87.8%.