Recognition of Food Ingredients—Dataset Analysis

Abstract

Nowadays, food waste is seen as a complex problem with effects on the social, economic, and environmental domains. Even though this view is widely held, it is frequently believed that individual acts have little to no impact on the issue. But just like with recycling, there may be a significant impact if people start adopting more sustainable eating habits. We suggest using a cutting-edge convolutional neural network (CNN) model to identify food in light of these factors. To improve performance, this model makes use of several strategies, such as fine-tuning and transfer learning. Additionally, we suggest using the Selenium library to create a dataset by employing the web scraping technique. This strategy solves the problem that many open-source datasets have with the overrepresentation of foods from the Asian continent by enabling the addition of foods to the dataset in a customized way. First, using the PRISMA methodology, a thorough examination of recent research in this field will be carried out. We will talk about the shortcomings of the most widely used dataset (Food-101), which prevent the ResNet-50 model from performing well. Using this information, a smartphone app that can identify food and suggest recipes based on the ingredients it finds could be developed. This would prevent food waste that results from the lack of imagination and patience of most people. The food recognition model used was the ResNet-50 convolutional neural network, which achieved 90% accuracy for the validation set and roughly 97% accuracy in training.