Improved food image recognition by leveraging deep learning and data-driven methods with an application to Central Asian Food Scene

Abstract

The burden of diet‑related diseases is high in Central Asia. In recent years, the field of food computing has gained prominence due to advancements in computer vision (CV) and the increasing use of smartphones and social media. These technologies provide promising potential in many applications by facilitating real‑time information retrieval from food images for efficient digital food journaling, smart restaurants, and supermarkets etc. Yet, to develop a robust CV model for food information retrieval, a large‑scale high quality dataset is required. Several food dataset have been developed covering Western, Mediterranean, Chinese etc. cuisines. These dataset solve the simpler problem of food classification with single food item per image, which is not practical for real‑life scenarios, where meals typically consist of multiple food items. To address this gap, we developed a large‑scale high‑quality Central Asian Food Scenes Dataset for food localization and detection. The dataset contains 21,306 images across 239 food categories, 69,856 instances. ed images. To evaluate the dataset, we performed the parametric experiments with the object detection models, with the best results achieved using YOLOv8xl (mAP50 score of 0.677).