Short answer
Packaging lines generate large reject archives: images of pouches, trays, caps, cartons and labels that an inline camera rejected, usually with a rejection reason. Seal and closure defects such as channel leaks, wrinkles, contamination in the seal and misaligned caps are visible in those images. Reasons are machine-assigned, reject images over-represent false calls, and few lines keep accepted images.
What a packaging line produces
A filling and packaging line runs fast, with inline cameras checking each package: seal integrity, cap and closure position, fill level, label presence and print, date code legibility, and foreign material. Rejected packages are diverted to a bin. Many systems save the rejected image with a reason code; fewer save accepted images, because storage was costly when the system was installed.
The result is an unusual asset. The rejects are plentiful, since a line may reject thousands of packages in a shift when a fault occurs, and they are tied to timestamps, a lot and a reason. But they carry the system's judgment, which may be wrong, and they represent an unbalanced sample of what the line makes. The broader pattern is covered in in-line machine vision archives. This page sits in the inspection cluster of the humanoid data section.
What the data can and cannot support
It can support seal, closure and package-integrity detection, anomaly detection when paired with defect-free images, rejection-reason classification, false-call filtering, and evaluation sets for new inspection hardware. Anomaly methods are usually benchmarked with defect-free training images and pixel-precise test annotations, as in the MVTec AD set of 15 categories (Bergmann et al., 2021), so a packaging archive is more useful when it includes normal images. The wider options for sourcing defective and normal data together are in sourcing defect image datasets.
It cannot show package handling. A humanoid picking up pouches or opening cartons needs grasp, force and deformation data that inspection images do not hold. It cannot reveal whether a hidden leak exists, since cameras see surfaces; leak tests, seal strength tests and pressure decay are separate measurements. And food, pharma or hazardous goods packaging adds regulatory context beyond appearance.
Defect families in packaging
| Family | Examples | Imaging considerations |
|---|---|---|
| Seal | Channel leaks, wrinkles, creases, product in seal, partial seal | Backlight or transmitted light often needed |
| Closure | Missing cap, cross-threaded, high cap, broken tamper band | Side views and top views |
| Fill and headspace | Under- or over-fill | Covered in fill-level and closure inspection |
| Container | Dents, cracks, scuffs, deformation | Reflective surfaces, glare |
| Label and print | Skew, missing, smeared, wrong lot code | Covered in label and print quality inspection |
| Foreign material | Fragments, hair, glass | Covered in foreign object detection |
| Case and pallet | Crushed cartons, open flaps | Different camera stage |
Reflective or transparent packaging deserves a specific check. MVTec AD 2 includes transparent and reflective objects and a split captured under varied lighting, which reflects that the benchmark authors treat those factors as harder cases (Heckler-Kram et al., 2025). Metallized film and clear trays are common sources of glare artifacts.
Illustrative reject-archive extract
Illustrative; values are synthetic.
| Field | Example |
|---|---|
| image_id | rej_0048213 |
| timestamp, line_id, station_id | 2026-05-14T03:22Z, line_7, st_2 |
| product_sku, lot_id | SKU_A31, L_9042 |
| machine_reason | seal_contamination |
| machine_score | 0.91 |
| operator_review | confirmed / false_call / not_reviewed |
| final_cause | film_temperature_drift |
| accepted_sample | Separate stream of 1-in-200 accepted images |
| recipe_version | R_2026_04 |
The operator_review and accepted_sample fields convert a reject pile into training data. Without them, the set is a list of machine opinions.
Quality criteria
False-call rate. Ask what share of rejects are good packages. High false-call rates are normal; they need to be labeled, not hidden.
Accepted images. Ask whether any exist and how they were sampled. A 1-in-N sample of accepted images across lots gives a representative normal class.
Recipe and lot structure. Packaging changes by SKU, film, tooling and season. Hold out lots or SKUs so a model does not memorize a design; see rare defects and class imbalance for why this matters with skewed classes.
Cause linkage. A reject spike usually traces to a process cause. Linking images to a documented cause makes them more informative than the reject reason alone.
Printed content. Packaging images contain brand names, ingredient text, barcodes and handling symbols. Standardized symbols, such as those in ISO 780 for marking distribution packages to convey handling instructions (ISO, 2015), are not sensitive, but brand artwork and lot codes are.
Preparing a packaging archive for use
Separate reject causes. A seal reject caused by film temperature drift looks different from one caused by a contaminated product stream. Cluster by cause so each class is coherent.
Balance time. Rejects arrive in bursts when a fault occurs, so a month of images may be dominated by two incidents. Sample across incidents and keep the burst structure in metadata, otherwise a model learns the incident.
Keep unreviewed images flagged. Images with no operator review are not negative examples. Label them unknown and exclude them from scoring.
Reproduce the camera view. Save the crop and mask that the inline system applied, since region-of-interest choices determine what the model can see. See imaging conditions and metadata.
Supplier types
Food, beverage, pharmaceutical, personal care and consumer goods manufacturers all run packaging lines, as do contract packagers. Packaging machinery makers and vision integrators hold test images from customer trials, though customers often own those. Ask who controls the archive: the plant, the brand owner or the integrator.
For natural products where appearance is graded rather than toleranced, see food quality grading images.
Rights and provenance
Package images show brand artwork and trademarks, and they can reveal unreleased packaging designs. A contract packager's images may belong to or be restricted by the brand owner. Ask for written confirmation of who can license, and whether the pre-launch packaging appears. Blurring text does not resolve ownership. This is not legal advice.
Limitations and extra preparation
Reject archives are shaped by the line's settings. A vision system with a tight threshold rejects many good packages; a loose one lets defects through, and the archive only contains what was flagged. Rejection reason codes are often assigned by the machine, not confirmed by a person, and they may be overwritten or reused when a line is reconfigured. Normal images are frequently not kept at all, which limits anomaly detection unless the line owner agrees to retain a sample. Transparent films, foils and printed artwork change from one product run to the next, so a model trained on one run's packaging may not transfer to the next design.
The data supports seal and package-integrity inspection and false-call filtering, not package handling: it records no grasp, force or deformation. Plan for extra preparation: confirming reason codes against a manual review sample, recording camera and lighting per period, masking brand artwork or lot codes where needed, and splitting by product run or date so that an evaluation actually tests unseen conditions.
Questions to ask line owners
- Which inspection system saved the images, and under which recipe versions?
- Are accepted images kept, and at what sampling rate?
- Is the reject reason machine-assigned only, or reviewed by an operator?
- What is the false-call rate for each reason code?
- How many distinct SKUs, films and closure types appear?
- Are leak or seal-strength test results linked to any images?
- Does the archive include unreleased or customer-owned packaging?
- For how long are images retained, and in what format and resolution?
Discuss your requirements
If you need packaging and seal defect imagery from US manufacturers and packagers, describe the package types, defect families 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.
Tell SourceX what your evaluation set needs to cover
Sources
- The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — Anomaly benchmarking with defect-free training images and pixel-precise test ground truth, 15 categories.
- ISO 780:2015 Packaging - Distribution packaging - Graphical symbols for handling and storage of packages — Distribution-package handling symbols as an example of printed marks inspected on packages.
- The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection — Benchmark scenarios with transparent or reflective objects and varied lighting.