Abstract
Background Assessment of dietary intakes is notoriously laborious and generates information that requires a lot of effort in coding and subsequent analysis. Furthermore, keeping track of the everyday diet via taking photos of the meals might enhance the accuracy of the dietary assessment and reduce reporting and recall biases. Currently, Computer Vision (CV), which is a subfield of Artificial Intelligence (AI) is being utilized for various applications in food recognition such as smart restaurants, supermarkets, and nutritional assessment to increase social awareness of a healthy lifestyle. Thus, information mined from food images using CV could have considerable potential in dietary interventions. Objectives The overarching project aims to apply CV techniques to identify main dietary factors in association with corresponding cardio-metabolic factors in Kazakhstan. In order to achieve this, we have to first develop a model for detecting and identifying food items unique to local Central Asian cuisine. We will then apply transfer learning from pre-trained food classification models to our custom dataset. Further, we will link the nutritional content to the food classes, such that the model will provide the assessment based on the longitudinal dietary patterns. Methods A Telegram Bot was created to collect food images unique to Central Asia as well as other dietary and lifestyle factors. For each food class, approximately 1,000 images are to be collected and annotated. In the case of rare food items, data augmentation techniques will be applied. Results To date, we have collected images for about 8 classes of foods and 2 classes of beverages unique to Kazakhstan. More than 4,000 images have been collected and annotated. While the rest of the classes are being pre-processed, we are now performing parametric experiments with EfficientNet and ResNet deep learning models. Further details will be provided during the presentation. Conclusions The creation of the Central Asia food datasets will help to better explore and examine the dietary patterns which will allow researchers to conduct both nutrition and dietary surveillance in a more effective manner.