Large Scale Visual Food Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Food Science

Background:

• Food recognition is crucial for human health and well-being, impacting food choice and intake.
• Existing food recognition datasets lack scale, hindering advancements in computer vision for the food domain.
• Progress in generic visual recognition has not been matched in specialized food-related tasks.

Purpose of the Study:

• Introduce Food2K, the largest food recognition dataset to date, featuring 2,000 categories and over 1 million images.
• Establish a new benchmark for developing advanced models in food visual representation learning.
• Propose and validate a novel deep progressive region enhancement network for improved food recognition.

Main Methods:

• Developed Food2K dataset, surpassing existing benchmarks in scale (categories and images).
• Proposed a deep progressive region enhancement network with progressive local feature learning and region feature enhancement using self-attention.
• Conducted extensive experiments on the Food2K dataset to evaluate the proposed method.

Main Results:

• The proposed deep progressive region enhancement network demonstrated effectiveness on the Food2K dataset.
• Achieved superior performance in food image recognition, retrieval, and cross-modal recipe retrieval.
• Verified enhanced generalization capabilities across diverse food-related computer vision tasks, including detection and segmentation.

Conclusions:

• Food2K serves as a challenging benchmark for large-scale, fine-grained visual recognition in the food domain.
• The proposed network advances food recognition and can serve as a backbone for various food-relevant tasks.
• Food2K is expected to drive progress in food-relevant AI applications, including nutritional understanding.