Hyperspectral Imaging × AI: The Hard Problem That Held Back Food Inspection — and How It's Being Solved

HSI Always Saw Too Much

Hyperspectral imaging (HSI) captures food across hundreds of wavelength bands simultaneously, embedding chemical information into every pixel — moisture, protein, fat, microbial load, foreign objects. In theory, it's the perfect inspection tool: non-destructive, fast, and comprehensive.

In practice, it spent decades being too hard to use.

Two Fundamental Challenges

1. Data Volume

Where a standard color image has three channels, a hyperspectral image has hundreds. Processing that data at production-line speeds — frames per second, across an entire facility — overwhelmed conventional algorithms.

2. Spectral Unmixing

This is the deeper problem. Every pixel in a hyperspectral image records not a single material's signature, but a convolved mixture of the optical properties of everything present at that point. Scan a piece of chicken, and the spectrum you get is muscle, fat, water, and connective tissue all folded together into one signal. A contaminant or pathogen buried within that matrix produces a faint signal — masked by the dominant spectra of the food itself.

Traditional approaches — Principal Component Analysis (PCA), Support Vector Machines (SVM) — could handle simple cases, but complex food matrices, trace contamination, and multi-ingredient products consistently pushed them to their limits. Unmixing the signal cleanly enough to make reliable safety decisions was the unsolved problem.

How Deep Learning Breaks the Unmixing Barrier

Convolutional neural networks (CNNs) and other deep learning architectures don't attempt to unmix spectra explicitly. Instead, they take the full mixed signal as input and learn — from thousands of labeled examples — what patterns within that complexity correspond to contamination, spoilage, or quality deviation.

This end-to-end learning approach is precisely what makes it suited to the unmixing problem: the model internalizes the relationship between the convolved signal and the outcome, without requiring a clean separation of components first. It performs best exactly where traditional methods fail — low-concentration contaminants, complex ingredient profiles, subtle spectral shifts.

Recent work published in 2025 demonstrates the state of the art: a CNN-BiGRU-Attention model achieved high-precision non-destructive quantification of vitamin C, soluble solids, and soluble protein in apples. Applied across meat, poultry, produce, grains, dairy, and chocolate, deep learning has crossed the threshold from research tool to production-ready technology.

Beyond Food Science: Built Around the Hard Problem

Montreal-based Beyond Food Science (beyond) is a company that grew directly out of engagement with these technical challenges.

Founded in 2016 at Cintech Agroalimentaire, one of North America's leading food research centers, the platform was built from the ground up around spectroscopy and HSI. Co-founder Dr. Michaela Skulinova brings an unusual lineage: a PhD in chemistry, postdoctoral work in X-ray optics at UCD Dublin, and Raman spectroscopy research at the Canadian Space Agency — where the core challenge is extracting meaningful signals from noisy, mixed spectral data in remote sensing instruments. That same skill set, transposed to food safety, maps directly onto the unmixing problem.

The platform is structured around three commercial modules:

beyond safety — Real-time pathogen and contaminant detection on the production line. AI generates contamination maps and triggers immediate alerts. The average U.S. food recall costs $10 million; early detection is the most cost-effective intervention.

beyond quality — Composition analysis (moisture, fat, protein), defect and foreign object detection, and AI-based shelf-life prediction. Quality-related costs run at 15–20% of revenue for many food manufacturers; catching deviations inline reduces both waste and rework.

beyond authenticity — Spectral fingerprinting to detect adulteration, ingredient substitution, and false origin claims. Global food fraud costs an estimated $15 billion annually.

Delivered as cloud-based SaaS with API integration, the platform scales from an Essential plan at $99/month through to enterprise deployment.

The Unmixing Problem in the Field

The real test is production, not the lab. Two recent deployments illustrate how far the technology has come.

ELROILAB deployed HSI + AI across a baby food production line and reported a 90% reduction in quality issues (October 2025). INNDEO commercialized detection of microplastics in minced meat using the INSPECTRA CHP® hyperspectral system (November 2025). The latter is particularly telling: microplastic spectra are among the most difficult to unmix from food backgrounds — their signal is weak and partially overlaps with organic matter. That this is now a solved production problem marks a genuine milestone.

Conclusion

HSI wasn't slow to reach the food industry because the concept was wrong. It was slow because the unmixing problem — extracting clean intelligence from convolved spectral signals — was genuinely hard. Deep learning has provided the key.

Beyond Food Science represents the practical translation of that technical breakthrough: a platform built by people who understand the spectral physics, commercialized with the rigor needed to meet food industry standards.

At flyeye, we connect the global innovators working at the intersection of optics, AI, and life sciences with the Japanese market. The food safety revolution is running ahead of schedule.

References:

• Beyond Food Science Official Website