Uncertainty-aware RGB and hyperspectral vision for food quality control

Overview

This system provides an inspection method that uses complementary imaging techniques to ensure consistent food product quality. It combines calibrated RGB and near-infrared/mid-infrared hyperspectral imaging through an AI-based segmentation and uncertainty-aware classification framework. The technology is designed to be used as an on-site inspection station that can quickly evaluate meat-based food products, with customers able to obtain quantitative measurements, annotated images, and a clear pass/review/fail decision.

Technical specifications

The technology integrates a controlled imaging chamber with a fixed-depth tray, diffuse and polarized lighting, and color calibration to minimize glare and drift. Key technical components include:

• THRESHOLD-based and geometric assessment: calibrated RGB imaging quantifies meat-piece size, shape, surface color, background appearance, and meat-to-gravy/jelly ratio.
• Spectral assessment: NIR/Mid-IR hyperspectral imaging provides additional spectral contrast to separate regions that are visually similar or affected by reflection or natural product variability.
• Uncertainty-aware machine learning: models are built around product-specific reference distributions, generating confidence scores, deviation flags, and a pass/review/fail decision.
• Measurement and decision output: within five minutes, the system returns annotated images, measured metrics, confidence scores, deviation flags, and a final decision.
• Deployment platform: models are managed through MLflow, allowing registry, version control, real-time inference, drift monitoring, and controlled retraining for consistent operation across products, lines, and sites.

Technology readiness level

The development is supported by substantial prior experience in hyperspectral sensing, chemometrics, and uncertainty-aware AI, but the fully integrated system is still in the proof-of-concept stage. Future validation steps include sample preparation, estimation of reference ranges, batch and data collection, model development, and deployment trials. Once validated, the selected configuration can be automated for at-line checks or continuous online monitoring during filling. Current readiness indicates a TRL of 3–4, with pilot validation already to begin immediately.