TLDR: This research introduces an AI-driven computer vision framework using semantic segmentation to estimate post-consumer food waste in university dining halls. By comparing before-and-after meal images, the system quantifies food consumption and waste for various dishes. The study found that U-Net and U-Net++ models, including their lightweight versions, effectively predict food waste, with lighter models offering faster processing. This approach provides a scalable, cost-effective, and non-intrusive method for monitoring food waste and informing sustainability strategies.
Food waste is a significant global challenge, impacting the environment, economy, and humanitarian efforts. Billions of tons of food are wasted annually, depleting vital resources like water, land, labor, and energy. Institutional food services, such as university dining halls, are major contributors to this problem but often lack effective systems for monitoring waste.
Traditional methods for measuring food waste, like physical weighing or staff estimates, are often time-consuming, prone to human error, and lack the detailed information needed for long-term analysis. This is where artificial intelligence (AI), particularly computer vision, offers a promising solution.
A recent study introduces a new framework that uses computer vision to estimate food waste after meals. This cost-effective system uses semantic segmentation, a technique that identifies and outlines different objects in an image, to analyze photos of plates taken before and after a meal. The goal is to quantify how much food was consumed and how much was left over.
The researchers trained four deep learning models, including U-Net, U-Net++, and their lighter versions, using images of five common Iranian dishes. These models were evaluated using various metrics, including a custom-designed ‘Distributional Pixel Accuracy’ (DPA) metric, which focuses on how well the models estimate the proportions of different food items rather than just pixel-perfect outlines. This DPA metric is particularly useful for food waste estimation, where knowing the quantity of leftovers is more important than the exact shape of every crumb.
The study found that all models performed well, with at least one model for each food type achieving or exceeding 90% DPA. This indicates a strong ability to estimate pixel-wise food proportions accurately. Interestingly, the lighter versions of the models, which have fewer parameters and require less computational power, performed comparably to or even better than their larger counterparts. These lightweight models also offered faster processing, achieving real-time speeds of over 80 images per second on an NVIDIA T4 GPU, making them practical for real-world deployment.
The research highlighted that dry and more rigid foods, like rice and fries, were segmented with higher accuracy. More complex or semi-liquid dishes, such as stews, presented greater challenges, especially in post-consumption images where food tends to spread or leave smeared traces. This difficulty arises from the ambiguity in defining precise boundaries for such foods, even for human annotators.
Despite some limitations, such as relying on 2D images (which don’t capture food volume or density) and a limited variety of food types, this framework represents a significant step forward. It offers a scalable, contactless, and cost-effective solution for continuous monitoring of food consumption in large-scale food service environments. The insights gained from such a system can help dining hall managers adjust serving portions, reduce waste, and implement more data-driven sustainability strategies.
Also Read:
- Smart Labeling: How ConformalSAM Improves Segmentation with Foundational Models
- Gaze-Guided Robots: Enhancing Efficiency and Robustness with Human-Inspired Vision
Future work could explore 3D analysis to capture volumetric information, hybrid approaches combining image analysis with weight measurements, or integrating advanced AI techniques like Type-2 Fuzzy Logic to better handle the visual uncertainty of semi-liquid foods. For more details, you can read the full research paper here.
