Inline visual inspection of meat products in gravy or jelly via deep learning and computer vision

Technology
In development
University

Computer vision system that automatically inspects meat products in gravy or jelly using deep learning segmentation to isolate individual meat pieces, then measures piece size, shape, color, and meat-to-background ratio. Designed for inline, at-line, or near-line deployment to provide repeatable, objective quality assessment without manual inspection.

Overview

This computer vision system provides automated visual quality inspection for meat products packaged in gravy or jelly. It uses a two-stage pipeline: a foundation-model segmentation approach isolates individual meat pieces within the surrounding matrix, even in complex scenes where pieces overlap or vary in morphology; then classical computer vision techniques extract quantitative attributes for each piece.

The system is designed for flexible deployment — inline during processing or transfer, at-line, or near-line sampling immediately after filling — delivering automated, repeatable visual quality assessment that reduces reliance on manual inspection and improves consistency.

Technical specifications

Key features:

  • Two-stage pipeline combining deep learning segmentation with classical computer vision measurement
  • Foundation-model segmentation (Meta SAM) achieves piece-level delineation even with overlapping pieces and variable morphology
  • Per-instance measurement of piece size and size distribution, shape descriptors, meat-piece color, background (gravy/jelly) color, and meat-to-background area ratio
  • Modular design: segmentation backbone can be upgraded independently of the measurement pipeline
  • Flexible deployment configurations: inline, at-line, or near-line
Technology readiness level

The segmentation approach has been validated in preliminary testing on images of meat in gravy, confirming that individual piece-level delineation is achievable. The measurement pipeline relies on standard computer vision methods and requires no additional deep learning.

Planned validation phases: algorithm development and tuning on a representative dataset; hardware integration and pilot on a production line, including camera mounting, lighting, and trigger timing; and final validation comparing automated measurements against manual reference assessments to document accuracy, repeatability, and false-flag rate.


About University of Manitoba

The University of Manitoba is a comprehensive public research university in Winnipeg with a broad portfolio spanning discovery to translation. Industry engages through an on-campus research and technology park that colocates companies with university labs, and through a health sciences campus embedded in the city’s hospital district for clinical studies and clinical-grade prototyping. A well-established co-op and internship ecosystem and the region’s logistics hub position make collaboration, field trials, and pilot-scale testing accessible. Research is supported by competitive funding from NSERC, CIHR, SSHRC, and the Canada Foundation for Innovation. A dedicated technology transfer office provides IP strategy, licensing, and startup support with flexible agreements for corporate partners.

Halo home
Partner smarter. Move faster.
Get new partnering requests
delivered to your inbox.