Retrieval and classification of food images

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Automatic food understanding from images is vital for health and economic applications, particularly in food intake monitoring.
• Current systems require robust retrieval and classification engines for diet monitoring, often integrated into wearable devices.
• Food image analysis presents challenges due to high variability and deformability of food items.

Purpose of the Study:

• To survey existing food image processing methods from early approaches to state-of-the-art.
• To investigate the role of image representation in food retrieval and classification engines.
• To propose and evaluate a novel image representation for enhanced food understanding.

Main Methods:

• A comprehensive survey of food image processing literature.
• Development and introduction of the UNICT-FD1200 dataset, featuring 4754 images of 1200 distinct dishes with labeled categories.
• Evaluation of various state-of-the-art image representations on the UNICT-FD1200 dataset.
• Proposal of a new Anti-Texton based image representation.

Main Results:

• The UNICT-FD1200 dataset captures real-world meal acquisitions with geometric and photometric variations.
• Performance tests revealed the limitations of existing representations for food retrieval and classification.
• The proposed Anti-Texton representation demonstrated superior performance in encoding spatial information, outperforming other methods.

Conclusions:

• Effective image representation is fundamental for accurate automatic food understanding.
• The Anti-Texton representation offers a promising approach for improving food retrieval and classification systems.
• This research contributes a valuable dataset and a novel method for advancing food image analysis.