Petia Radeva (AIBA/UB) — Boosting the limits of Data-centric Deep Learning for Visual Food Computing

/ Thu, 03/20/2025 - 10:00

Don't miss any Success Story following us on X and LinkedIn!
@RES_HPC RES - Red Española de Supercomputación @res-icts.bsky.social

Check this Success Story at our LinkedIn: Boosting the limits of Data-centric Deep Learning for Visual Food Computing

💡 A new success story about identifying food nutrients with images! 💡

📋 "Boosting the limits of Data-centric Deep Learning for Visual Food Computing" led by Petia Radeva from the AIBA Consolidated Research Group within Universitat de Barcelona

Visual Food Recognition uses Deep Learning (DL) and Computer Vision algorithms to automatically extract food information from a single image, providing a natural, objective, precise, and easy-to-use tool for identifying food items, digitally storing them, and converting them into estimated macro and micro-nutritional information.

However, the data needed to train models of this kind must be of excellent quality and be thoroughly categorized and prepared, due to the fine-grained distinctions between foods, their high intra-class variability, and long-tailed category distributions. To overcome these challenges, large training datasets are necessary, making HPC resources crucial to develop these models.

🖥️ Thanks to RES supercomputer MareNostrum5 ACC from Barcelona Supercomputing Center, the team addressed these complex challenges and integrated VLMs and LLMs, being able to extract detailed ingredient information from a single dish image.

The resulting models from this training achieve State-of-The-Art (SoTA) performances on several popular food benchmarks, improving over 2% the performance compared to other SoTA MODELS. Additionally, their results have been published in reputed congresses, making them accessible and bringing AI-food applications closer to real-world deployment.

📸 The images show the challenges and advancements in fine-grained food classification present in the project, showcasing comparisons of similar dishes in Fig1, detailing the methodologies employed in Fig2, and demonstrating the effectiveness of LLM-guided systems for accurately predicting food classes.