USDA reference-image ranking
Higher Spearman ρ indicates closer agreement with the known order of eight reference photographs.
An ordering result is not an official grading accuracy estimate.
REPRODUCIBLE RESEARCH • CHAPTER 6
From an image of beef to an auditable decision.
Segment meat and fat, quantify marbling within the eye-muscle region, and combine the image with appropriately sourced carcass measurements. Each result carries its assumptions and limitations.
A traceable chain, not a black-box grade
01Detect
02Segment
03Measure
04Verify
05Decide
The normative rules may refuse classification when required evidence is absent or unverified.
RESEARCH EVIDENCE
Numbers loaded from local study outputsWhat the reference experiments actually show
Loading study results…
n = 8
A measurable, explainable pipeline
01
Anatomically scoped marblingGrid values within the eye-muscle ROI, rather than treating outer fat as marbling.
02
Evidence-aware standardsImage estimates are kept distinct from instrument measurements and reference labels.
03
Boundary-aware decisionsThe fuzzy layer indicates review near thresholds without changing crisp standards.
METHODOLOGY
The five processing stages
Scientific boundary: the included USDA and GOST images are reference illustrations. Experiments on synthetic carcasses or line frames do not establish performance on independent production animals. Full reproduction of dissertation §6.8 remains incomplete.
INTERACTIVE RESEARCH WORKFLOW
Analyze an image, inspect every stage.
Choose a reference sample or upload your own image. The default method requires no neural-network checkpoint.
ANALYSIS TRACE
What the algorithm found
Stages 1–3 · Visual segmentation and ROI
Inspect the original, segmentation overlay, and the per-cell grid. Individual masks are available below.
Grid index explained
The index is the unweighted mean of eligible cells. Gray cells are excluded under the fixed ROI coverage threshold; this differs from the total fat percentage in the entire meat cut.
Stages 4–5 · Evidence-aware decision support
The original crisp standards remain authoritative; fuzzy membership flags ambiguous boundary cases.
PRODUCTION-LINE RESEARCH
Track cuts. Verify evidence. Recommend the next action.
Use fixed-camera conveyor video with an initially empty belt. Review track IDs before attaching measurements. These are advisory decisions, not machinery commands or certified carcass grades.
01Capture video
02Detect and track
03Check image suitability
04Associate measurements
05Review recommendations
A full beef cut on a conveyor is not necessarily a suitable anatomical cross-section for USDA/GOST marbling assessment. The interface requests cross-section verification and withholds unsupported operational grading.
Production-line video and evidence
Run image-only first to obtain the video SHA-256 and track IDs. Use those exact identifiers to associate external measurements during a second pass.
REPRODUCTION & TRANSPARENCY
Research results, with their limits visible.
All numerical comparisons below are read from the study outputs supplied with this application. They distinguish reruns, synthetic scenarios, unavailable experiments, and biological validation.
Reference-image comparison
USDA n = 8, GOST n = 4. Spearman correlation with predefined reference order; LOOCV error is expressed in USDA score units.
NBQA-2022 simulation
Independent synthetic marginals: tests USDA grading-rule behavior, not performance on real carcasses.
Thesis agreement audit
Computed results versus stated Chapter 6 values. Matching numbers do not establish independent scientific validation.
Image perturbation sensitivity
Eleven controlled image changes. Spearman ρ across the same eight USDA references; this does not test independent biological samples.
Fuzzy boundary review
For k = 2 and nominal measurement noise (σ × 1), compare referral rate with the proportion of noise-induced errors flagged.
Fresh IMARS pilot: evidence, not a replacement for Chapter 6
Twelve reference images reprocessed with the unchanged classical pipeline; a controlled synthetic production-line pilot. Independent real-carcass validation is not claimed.
—
Adaptive border recovery v2 · controlled comparison
Compare frozen, v1 and v2 masks. The new pixels are hypotheses, not expert-labelled tissue.
Loading…
Whole-cut recovery and dynamic grid · new experiment
Compare the original cut and predicted fat/lean masks with an exploratory pale-fat boundary recovery. Study outputs preserve the frozen grading pipeline.
Loading results…
Evidence-first decisions · tested safety resolution
Selection-aware grid comparison and a conservative review policy. Image-only estimates are not released automatically.
Loading results…
Grid selection and selective review · controlled experiment
Five grid resolutions on the same 12 references; independent simulated development/test halves for uncertainty-based review. This is not new carcass validation.
Loading results…
Scientific evidence dossier · fresh analysis
Reference masks, leave-one-out calibration, GOST limitations, and the original 32-cut synthetic replay. Each figure is generated from data and is not independent carcass validation.
—
Reproduction claim-by-claim
Sources, assumptions and incomplete work
COMPATIBILITY WITH CHAPTERS 1–5
Optional legacy image models.
The Chapter 6 classical pipeline remains the default. Older image-state classification and H5 tissue segmentation are opt-in, separate research extensions.
Visual-state classifier
Optional YOLO classifier from the early research chapters. “Fresh” and “Spoiled” are dataset appearance labels—not microbiological safety or permission to sell meat.
How to enable safely
Install the optional dependencies, review the original model's class-name mapping and SHA-256 in its config, and explicitly enable the extension using TRUSTEDBEEF_ENABLE_LEGACY_FRESHNESS=1. The model is never loaded when the default app starts.
H5 fat/lean segmentation
Optional alternative for Stage 2. The default CIE Lab K-means remains unchanged; an H5 mask may not be biologically validated and can be unsuitable for grading.
How to enable safely
Supply yolo_meat_segmentation_model.h5 and the validated preprocessing/class mapping contract in models/. Checkpoint diagnostics may flag near-random initialization; that output must remain exploratory.
Legacy modules are provided for historical compatibility only. They do not expand the scientific claims established by the Chapter 6 reference and synthetic studies.