AI for Diet Quality Assessment from Lebanese Food Images Using the Diet Quality Questionnaire
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Traditional methods for assessing diet quality depend on people reporting what they eat and experts classifying the information. This process takes considerable time and effort which makes it difficult to use on a large scale. This thesis looks at how Lebanese food images can be classified automatically using the Diet Quality Questionnaire (DQQ). For this study, a dataset of 10,672 food images was collected from the web, it includes 230 food categories and 28 DQQ labels. For each food item, the local nutritionists indicated whether or not it belonged to each of the 28 DQQ categories. We used transfer learning to train a Direct DQQ Classifier to predict the DQQ categories directly from food images. In addition, the Food Item Classifier was used to identify the food item first, and then we mapped it to the corresponding DQQ categories using the annotations provided by the nutritionists. Furthermore, GPT4.1 and GPT-5.5 Medium were also tested, but they were not trained on this task beforehand. On the web image test set, the Food Item Classifier Pipeline gave the best results, with a subset accuracy of 0.8013 and a micro F1 score of 0.8801. The Direct DQQ Classifier reached 0.5631 and 0.8008, respectively. We also tested the models on 226 wearable camera images. For these images, GPT-4.1 had the best results, with a subset accuracy of 0.3363 and a micro F1 score of 0.6713. GPT-5.5 Medium reached 0.3097 and 0.6447. The Direct DQQ Classifier had a micro F1 score of 0.5439, which was higher than the 0.4812 reached by the Food Item Classifier Pipeline. These results show that food images can be useful for DQQ assessment. However, the results were affected by class imbalance, foods that look similar, fewer examples in some categories and the variety of meals found in everyday settings.