DeshiFoodBD: Development of a Bangladeshi Traditional Food Image Dataset and Recognition Model Using Inception V3

Abstract

For a multitude of reasons, including restaurant selection, travel location selection, dietary calorie intake, and cultural awareness, traditional Bangladeshi cuisine picture classification has become more important. However, creating an efficient and usable traditional labeled (English and Bengali) food dataset for study in Bangladesh is quite difficult. In this article, the ‘DeshiFoodBD’ dataset is given in both Bengali and English for traditional Bangladeshi food classification. Web scraping and camera photos (digital, smartphone) are used to create food images. The dataset includes 5425-labeled pictures of 19 popular Bangladeshi dishes, including Biriyani, Kalavuna, Roshgolla, Hilsha fish, Nehari, and others. There are a number of convolutional neural network (CNN) architectures that may be utilized with this dataset including ImageNet, ResNet50, VGG-16, R-CNN, YOLO, DPM, and so on. It is currently available at Mendeley data repository for research purposes and further use. Our proposed Inception v3-based food recognition model, DeshiFoodBD-Net, had higher test accuracy of ~97% with the DeshiFoodBD dataset.