Short answer
Food quality grading images show produce, meat, baked goods or ingredients labeled with a grade, defect type or accept decision, usually by trained graders against a published or customer standard. Natural products vary by season, variety and lot, so grades are a judgment against a standard rather than a tolerance. Sorter archives are a common source; they rarely include handling data, and grader disagreement is part of the data.
Grading is a judgment, not a measurement
Machined parts are inspected against dimensions and tolerances. Food is graded against descriptions: color range, size class, blemish area, shape, firmness, maturity, foreign matter. A grade standard may combine numeric limits (for example a blemish area) with descriptive terms that graders interpret; check the applicable standard and version for how each criterion is defined. Customers may add their own specifications on top.
For a data buyer this has three consequences. First, labels carry the grader's and the standard's identity. Second, natural variation across varieties, regions and seasons is large, so a dataset from one harvest may not describe another. Third, borderline items are common and informative; see borderline samples.
This page covers sourcing grading images. It sits in the inspection cluster of the humanoid data section. For recognizing food items rather than grading them, see food item recognition.
What grading images can and cannot support
They can support grade classification, defect detection (bruises, rot, mold, scars, discoloration), size and shape estimation, ripeness estimates from color, sorting decisions and evaluation of sorting equipment. Sorter systems generate large volumes with labels from the machine's decision or a reviewer's correction.
They cannot show how to handle the product. Picking a ripe tomato without bruising it, bagging leafy greens or placing a pastry needs grasp, force and deformation data; images of graded items give none. Datasets of people working with food, such as the 100-hour EPIC-KITCHENS-100 egocentric kitchen video collection (Damen et al., 2020), record kitchen activity and carry no quality-grade labels, so they are a different asset for a different purpose.
Images also miss non-visual quality: sugar content, internal defects, texture, odor and safety attributes such as pathogens. Hyperspectral, near-infrared or X-ray systems capture some of these; the labels there come from lab measurements, not appearance.
Where grading data is generated
| Setting | What is captured | Label source |
|---|---|---|
| Packhouse sorters | Multi-camera images per fruit or item, with machine grade | Sorter decision, sometimes graders' corrections |
| Processing line inspection | Images of cut, cooked or formed products | Inline system, QA sampling |
| Receiving inspection | Photos of incoming lots and rejected loads | Inspector notes, grade sheets |
| Meat and seafood grading | Carcass or fillet images, trim and defect marks | Graders |
| Bakery and snack lines | Product shape, color, size and breakage | Line cameras, QA |
| Third-party audits | Photos with findings | Auditor reports |
Existing archives are the main route. Because harvests are seasonal, an archive's value depends on how many seasons it spans.
Illustrative package specification
Illustrative; values are synthetic.
| Element | Example |
|---|---|
| Commodity and variety | Apples, variety_12, origin region region_B |
| Season and lot | Harvest year, storage duration, lot IDs |
| Images | Four views per item on a rotating conveyor, uniform light, 8-bit color |
| Grade label | Grade code from named standard and version, plus customer class |
| Defect label | Type, location, area estimate |
| Grader information | Grader ID (pseudonymous), experience band, duplicate-grade subset |
| Measurements | Size, weight, firmness or sugar from instruments for a subset |
| Splits | By season and lot; hold out one season |
| Reference | Color chart or gray card in each image |
Quality and acceptance criteria
Named standard and version. A grade label has no meaning without the standard that defined it. Require the document name and revision, plus any customer override.
Grader agreement. Request the grader disagreement rate on a duplicate-labeled subset; see inter-inspector agreement. If graders often disagree on borderline items, the label noise is a property of the task, and the evaluation must account for it.
Season coverage. Seasonality changes color, size and defect prevalence. Hold out entire seasons and varieties to measure generalization; see domain shift across lines.
Lighting and color control. Color-based grades require stable illumination and a color reference. Check for lamp aging and camera white balance drift.
Sorter bias. If the label is the sorter's decision, models learn to copy it. Prefer grader-reviewed labels for evaluation.
Foreign material. Stones, plastic and leaves appear in raw product streams; they may be labeled separately. See foreign object detection.
Sourcing routes: existing archives versus a planned capture
Existing sorter and line archives are usually the fastest route and carry real production variation, but their grade labels reflect whichever standard and machine settings applied at the time. They are also uneven: some seasons are well covered and others have been overwritten.
A planned capture can fill gaps, for example imaging a graded set of items across the full grade range under controlled light with a duplicate-graded subset. That requires a cooperating packhouse or processor, access to a harvest window and graders' time, so treat it as a scoping discussion. Seasonal timing means a missed window may cost a year.
For fresh produce, a calibrated reference set photographed alongside each batch makes cross-season comparison far easier, and costs little compared with relabeling afterwards.
Preparation the data holder needs to do
Group images by item, because multi-view captures of the same fruit are correlated. Remove grower, buyer and brand identifiers from filenames and sidecar files. Reconcile grade codes across seasons when standards changed. Record the camera and light recipe per period. Where only machine grades exist, arrange a grader review of a stratified sample, concentrating on borderline and rare-defect items rather than random draws.
Packaged goods in the same plant raise different questions; see packaging and seal defects.
Rights and sensitivity
Food images do not usually contain personal data, but they can reveal grower identity through tags and crates, supplier quality through reject rates, and customer specifications. Grading specifications from retailers are often confidential contract terms. Ask who owns the images, whether grower or customer agreements restrict disclosure, and whether the license covers derived models. Anonymizing filenames does not resolve contractual limits. This is not legal advice.
Limitations and extra preparation
Grading images record a judgment made under a specific standard, by specific graders or machines, during a specific season. Varieties, growing regions and harvest conditions change appearance enough that a model trained on one season's archive may not hold up on the next, and the grade boundaries themselves may have changed between seasons; check the applicable standard and version for each period. Machine-sorter labels reflect sorter settings, which were tuned for throughput, not for training data quality. Grader disagreement on borderline items is real and should be measured rather than averaged away.
These images show the product's surface from a fixed camera; they carry no firmness, internal quality or handling information, and they cannot teach a robot to pick or pack produce without bruising it, which needs grasp and force data from robots or instrumented capture. Plan for extra preparation: reconciling grade codes across seasons, removing grower and buyer identifiers, recording camera and light recipes per period, arranging a grader review of a stratified sample, and splitting evaluation data by season or lot.
Questions to ask suppliers
- Which commodities, varieties and seasons do the images cover?
- Which grade standard and version were used, and what customer overrides apply?
- Are labels from the sorter, from graders, or from both?
- What is the grader disagreement rate on borderline items?
- What camera, lighting and color-reference setup produced the images, and did it change?
- Can images be grouped by item and lot for splits?
- Are lab measurements available for any subset?
- Do grower or customer contracts limit licensing of the images?
Discuss your requirements
If you need food or produce grading images from US packhouses and processors, describe the commodity, grade standard and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match.
Discuss your humanoid training-data requirements with SourceX
Sources
- Rescaling Egocentric Vision — Contrast case: EPIC-KITCHENS-100 egocentric kitchen video records food handling, not quality grades.