Transfer learning approach with EfficientNet to enhance food recognition systems for health monitoring

Abstract

Food image classification is a challenging problem with significant potential benefits for real-world applications such as nutritional and energy estimation. Most prior research has proposed various Convolutional Neural Network (CNN) architectures to tackle this issue. However, given the large size and diverse nature of food image datasets, there remains considerable room for improvement, particularly in terms of accuracy and training speed. Typically, neural networks trained on small image classification datasets benefit from using pre-trained weights from large-scale image classification datasets like ImageNet. In this study, we explore the balance between using pre-trained networks as feature extractors and fine-tuning networks for food image classification. By leveraging transfer learning with EfficientNetV2B0, we achieve higher accuracy in food image classification. On the largest publicly available food image dataset, FOOD-101, our proposed method improves the previous best accuracy from 77.40% to 81.62%, while maintaining a prediction speed of 23 ms on a GPU.