Abstract:
This paper proposes a machine-learning approach that leverages deep-learning techniques specifically using the YOLOv8 model for real-time food item detection from images and recommendations, helping pregnant women with diabetes manage their diets effectively. As a dataset, 5,016 raw images with 18 different food classes were collected. All the images were captured from different sources with various backgrounds and lighting conditions. Several preprocessing techniques were also used such as annotation, augmentation, etc. to prepare the dataset. By using online annotating tools several bounding boxes were drawn around the food items. To make the balance between food classes and increase the dataset size augmentation techniques were also performed. With the prepared dataset two versions (YOLOv5 and YOLOv8) of the YOLO family are used. Both of the models performed well and were able to detect objects on food images accurately. But comparatively, the YOLOv8 model achieved the highest mean Average Precision (mAP). The model achieved a 0.882 mAP50 score and a 0.709 mAP50-90 score which indicates that our model is capable of detecting food items from images pretty well. As our proposed method has a recommendation system this model was later used on a mobile application called mDiet. In our mDiet application, a recommendation system has been integrated that takes user health and pregnancy-related information as input, which further relates to detected food class, because the model is also integrated to detect food items from images by capturing food images using a mobile camera or selecting images from their device. By providing their information and food images pregnant women could benefit from the mDiet application in diet management and get useful personalized recommendations. Overall using the YOLOv8 model and the mDiet application we got our expected result.
