In this paper, we propose a recipe recommendation method that can heal both the provider and the recipient of food. In particular, we focus on the motivation of recipe contributors, and propose a recipe recommendation method that can heal both parties by considering not only the recipe procedure but also the relationship between the provider and the recipient, and their health conditions. Specifically, we extract feature words by learning from the items in the recipe data that contain a lot of information other than the cooking procedure, such as "title," "introduction of the recipe," "one-point information," and "trigger." In addition, each item of the recipe is analyzed for sentiment, and ranked based on the extracted feature words and sentiment values of the physical condition and relationship. We validate the usefulness of the proposed method by evaluating the relationship and physical condition feature words extracted by the proposed method and by evaluating the recipe ranking using Rakuten's recipe data.
A coronavirus pandemic is forcing people to be "at home" all over the world. In a life of hardly ever going out, we would have realized how the food we eat affects our bodies. What can we do to know our food more and control it better? To give us a clue, we are trying to build a World Food Atlas (WFA) that collects all the knowledge about food in the world. In this paper, we present two of our trials. The first is the Food Knowledge Graph (FKG), which is a graphical representation of knowledge about food and ingredient relationships derived from recipes and food nutrition data. The second is the FoodLog Athl and the RecipeLog that are applications for collecting people's detailed records about food habit. We also discuss several problems that we try to solve to build the WFA by integrating these two ideas.
Food image recognition tasks are generally addressed by using a closed dataset. In a real-world setting, however, the dataset is updated as the new class of food appears, and it is impossible to train a model that distinguishes all kinds of food in advance. Also, inter-class similarity and intra-class diversity make food image recognition tasks more challenging. Previous works have shown that a personalized classifier using an individual user's food records can deal with some of these problems. On top of that, we propose a personalized classifier that learns from not only the accumulating food records of the individual user but also the growing records of the entire user. To conduct a realistic experiment, we build a new dataset of daily food images using a food-logging application called FoodLog Athl. As a result, our proposed method significantly outperforms prior personalized classification methods for food image recognition in a realistic setting.
One of the major tasks in food computing is vision-based food calorie estimation. However, unfortunately, food image datasets annotated with calorie amounts are very hard to obtain. In fact, no large-scale food datasets annotated with calorie amounts exits as long as we know. However, we can see some Web sites which provide photos of food set menus with only total calorie values. Then, in this work, we crawl such data from the Web and use them as training data of a vision-based food calorie estimation model. To estimate calorie amounts of food items, the calorie values of each food item in meal photos should be known in general. However, they are not available in this setting. Then, we propose a model employing food segmentation which can estimate calorie amounts of each food item from total calorie values of set meal photos. The experimental results showed that our region-segmentation-based calorie estimation model was able to estimate calorie amounts of individual food items roughly.
With the popularity of health management applications, awareness of dietary management is increasing. When calculating the number of calories in a dish, discriminating between food regions is an important factor. However, when using deep learning, a large amount of data is required for training, and it is impractical to collect data for countless food categories. In recent years, a method called few-shot segmentation has been studied to learn a semantic segmentation model using a small amount of training data. In this study, we propose a few-shot and zero-shot segmentation model which targets food images to overcome the insufficient amount of food training data and show the effectiveness of the proposed model on a semantic segmentation task for new food classes. In the proposed model, we employed the word embedding pretrained with a large-scale recipe text dataset, which results in better accuracy than the previous methods.
