Chrintz-Gath Gustav, Daivadanam Meena, Matta Laran, McKeever Steve
Department of Informatics and Media, Uppsala University, Box 513, Uppsala, 75120, Sweden, 46 709901474.
Global Health and Migration Unit, Department of Women's and Children's Health, Uppsala University, Uppsala, Sweden.
JMIR Form Res. 2025 Aug 14;9:e70124. doi: 10.2196/70124.
Recent advances in computer vision, particularly in deep learning, have significantly enhanced object recognition capabilities in images. Among these, real-time object detection frameworks such as You Only Look Once (YOLO) have shown promise across various domains. This study explores the application of YOLO-based object detection for food identification and portion estimation, with a focus on its alignment with the Swedish plate model recommended by the National Food Agency.
The primary aim of this study is to evaluate and compare the performance of 3 YOLO variants (YOLOv7, YOLOv8, and YOLOv9) in detecting individual food components and estimating their relative proportions within images, based on public health dietary guidelines.
A custom dataset comprising 3707 annotated food images spanning 42 food classes was developed for this study. A series of preprocessing and data augmentation techniques were applied to enhance dataset quality and improve model generalization. The models were evaluated using standard metrics, including precision, recall, mean average precision, and F1-score.
Among the evaluated models, YOLOv8 outperformed YOLOv7 and YOLOv9 in both peak precision and F1-scores. It achieved a peak precision of 82.4%, compared with 73.34% for YOLOv7 and 80.11% for YOLOv9, indicating superior accuracy in both food classification and portion estimation tasks. YOLOv8 also demonstrated higher confidence in its predictions. However, all models faced challenges in distinguishing visually similar food items, underscoring the complexity of fine-grained food recognition.
While YOLO-based models, particularly YOLOv8, show strong potential for food and portion recognition aligned with dietary models, further refinement is needed. Improvements in model architecture and greater diversity in training data are essential before these systems can be reliably deployed in health and dietary monitoring applications.
计算机视觉领域的最新进展，特别是深度学习方面的进展，显著提高了图像中的目标识别能力。其中，诸如“你只看一次”（YOLO）这样的实时目标检测框架在各个领域都展现出了潜力。本研究探索基于YOLO的目标检测在食物识别和份量估计中的应用，重点关注其与国家食品局推荐的瑞典餐盘模型的契合度。
本研究的主要目的是根据公共卫生饮食指南，评估和比较3种YOLO变体（YOLOv7、YOLOv8和YOLOv9）在检测图像中单个食物成分并估计其相对比例方面的性能。
本研究开发了一个包含3707张标注食物图像、涵盖42种食物类别的自定义数据集。应用了一系列预处理和数据增强技术来提高数据集质量并改善模型泛化能力。使用标准指标对模型进行评估，包括精度、召回率、平均精度均值和F1分数。
在评估的模型中，YOLOv8在峰值精度和F1分数方面均优于YOLOv7和YOLOv9。它实现了82.4%的峰值精度，相比之下，YOLOv7为73.34%，YOLOv9为80.11%，这表明在食物分类和份量估计任务中具有更高的准确性。YOLOv8在其预测中也表现出更高的置信度。然而，所有模型在区分视觉上相似的食物项目时都面临挑战，这凸显了细粒度食物识别的复杂性。
虽然基于YOLO的模型，特别是YOLOv8，在与饮食模型对齐的食物和份量识别方面显示出强大潜力，但仍需要进一步改进。在这些系统能够可靠地部署到健康和饮食监测应用之前，模型架构的改进以及训练数据的更大多样性至关重要。
