Journal: Information Fusion
Year: 2023
Citations: 263
DOI: 10.1016/j.inffus.2023.101859

Abstract

Food category recognition has become increasingly important in various applications including dietary monitoring, nutritional analysis, and automated food service systems. Deep learning approaches have shown remarkable success in computer vision tasks, making them particularly suitable for food recognition challenges. This comprehensive review examines the application of deep learning techniques to food category recognition, analyzing various neural network architectures, datasets, and performance metrics. We discuss convolutional neural networks, transfer learning approaches, and advanced architectures including attention mechanisms and transformer models. The paper also addresses challenges specific to food recognition such as intra-class variation, inter-class similarity, and cultural diversity in food presentation. We examine current datasets, evaluation methodologies, and emerging trends in the field.

Summary

This comprehensive review examines the application of deep learning techniques to food category recognition, addressing an increasingly important area with applications spanning dietary monitoring, nutritional analysis, and automated food service systems. The research analyzes how deep learning approaches, particularly convolutional neural networks, have revolutionized computer vision tasks and proven especially suitable for the complex challenges of food recognition. The study provides detailed examination of various neural network architectures, from traditional CNNs to advanced models incorporating attention mechanisms and transformer architectures that have shown remarkable performance improvements.

The paper addresses unique challenges specific to food recognition that distinguish it from general object recognition tasks, including significant intra-class variation where the same food can appear very different depending on preparation methods, cooking styles, and presentation approaches. The research also examines inter-class similarity challenges where different foods may appear visually similar, and cultural diversity factors that affect how foods are prepared and presented across different regions and cuisines. These challenges require specialized approaches and careful consideration of training data diversity and model architecture design.

The study provides comprehensive analysis of current datasets used for food recognition research, evaluation methodologies for assessing model performance, and emerging trends that are shaping the future of the field. The research examines transfer learning approaches that leverage pre-trained models to improve performance with limited food-specific training data, and discusses how attention mechanisms and transformer models are being adapted for food recognition tasks. The review provides valuable insights for researchers and practitioners working to develop more accurate and robust food recognition systems for real-world applications.

Main Takeaways

• Complex Recognition Challenges: Food category recognition presents unique challenges including significant intra-class variation, inter-class similarity, and cultural diversity in food preparation and presentation that distinguish it from general object recognition tasks.

• Advanced Deep Learning Architectures: The field has evolved from traditional convolutional neural networks to sophisticated approaches incorporating attention mechanisms and transformer models, showing remarkable performance improvements.

• Transfer Learning Effectiveness: Transfer learning approaches that leverage pre-trained models and fine-tune them for food-specific tasks have proven particularly effective for improving performance with limited domain-specific training data.

• Diverse Application Portfolio: The technology enables important applications including dietary monitoring, nutritional analysis, automated food service systems, and food safety and quality control across various sectors.

• Dataset and Evaluation Challenges: Success requires diverse datasets that capture real-world food preparation and presentation variations, along with evaluation methodologies that accurately reflect practical performance requirements.

• Cultural Adaptation Requirements: Effective food recognition systems must account for cultural diversity in food preparation and presentation styles, requiring comprehensive training data and robust model architectures.