Food Recognition: A New Dataset, Experiments, and Results.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Nutrition Informatics

Background:

• Dietary monitoring applications require accurate food recognition from images.
• Existing datasets may not fully represent real-world scenarios like canteen trays with multiple food items.

Purpose of the Study:

• To introduce a novel dataset of canteen tray images for evaluating food recognition algorithms.
• To provide a benchmark framework for assessing the performance of these algorithms.

Main Methods:

• The dataset comprises 1027 canteen tray images with 3616 food instances across 73 classes, manually segmented.
• An automatic tray analysis pipeline was developed, involving region detection and food class prediction.
• Convolutional neural network-based features and three classification strategies were experimented with.

Main Results:

• The developed pipeline achieved approximately 79% accuracy in food and tray recognition.
• Performance was evaluated using various visual descriptors and classification approaches.

Conclusions:

• The proposed dataset and benchmark framework are valuable resources for advancing food recognition in dietary monitoring.
• The results demonstrate the effectiveness of deep learning approaches for this task.