How to flag packaging surface defects without uploading every production image

How can a surface inspection preserve context?

Start with a defined question such as whether a printed carton has a visible tear, stain or missing print region. Do not combine cosmetic scoring, barcode verification and seal integrity into one unqualified “pass” score. Each has different evidence and failure costs.

Use a trigger tied to the item rather than sampling a free-running stream and hoping every package is seen. Record the lot, line, recipe, camera settings and model version with each result. A missing or blurred image should produce an explicit review condition, not a default pass.

Which model should be evaluated?

Anomalib documents PaDiM as a patch-distribution approach to anomaly detection and localization. It is a candidate when the team can collect representative acceptable products, but a statistical anomaly is not automatically a commercially unacceptable defect. Variation in artwork, material, light and alignment must be included in evaluation.

Compare an anomaly model with a simpler rule-based baseline and, where labels support it, a supervised model. Inspect failures at full image resolution. A small crop can hide a tiny defect, and an attractive heatmap is not a release criterion. Define what happens when a label or image is ambiguous before results go to production staff.

What storage could a pilot require?

Estimate assumptions: one camera using the documented 1920 × 1200 resolution, five triggered packages per second, one frame per package, eight operating hours and an eight-bit grayscale stored working image. Resolution is a camera specification; the selected rate and operating time are workload assumptions, not a claimed camera configuration or inference benchmark.

Daily images = 5 × 8 × 3,600 = 144,000. Uncompressed grayscale bytes per image = 1920 × 1200 × 1 = 2,304,000. Saving all working images would therefore consume 331.776 GB per day, before metadata, replication and encoding overhead.

For a pilot, suppose the retention policy keeps 2% flagged examples plus a non-overlapping 1% random audit sample. Retained images would be 4,320 and the image subtotal 9.95328 GB per day. These percentages are chosen policy assumptions, not a predicted defect prevalence or detection rate. An increased reject rate changes retained volume; provision overflow handling instead of deleting evidence silently.

At five packages per second, items arrive every 200 milliseconds. That is the arrival interval, not an automatic latency allowance: a buffered or parallel pipeline can have different throughput and response-time limits. Determine the actual deadline from conveyor geometry and the disposition mechanism. Measure dropped triggers, queue age and the full capture-to-decision time under worst expected load.

What belongs in the pilot system?

Use one industrial camera, lens and fixed lighting enclosure; a trigger and lot identifier; a customer-operated edge computer with local buffer storage; an evaluated preprocessing/model pipeline; and a review workstation. A Jetson Orin NX-based industrial computer is one possible accelerator option, but the hardware, operating system and runtime combination needs testing. Add cloud object storage and the HyperLake environment only after deciding which images may leave the site.

Illustrative compute-module estimate, not a complete inference cost or quote: assume one NVIDIA Jetson Orin NX 8 GB module assigned to this camera, amortized over three years, a 40 W module power mode, eight operating hours per day and 250 operating days per year. NVIDIA lists a $649 USD volume suggested price at 1,000 units and describes the 40 W mode for Orin NX production modules; the price is for the module, not a full industrial computer. The U.S. Energy Information Administration reports a 2025 industrial electricity average of 8.51 cents per kWh; use the facility's tariff for a local estimate.

Module allocation = $649 ÷ 3 = $216.33 per camera-year. Assumed module electricity = 0.040 kW × 8 hours/day × 250 days/year = 80 kWh per year, or 80 × $0.0851 = $6.81 per camera-year. The subtotal is $223.14 per camera-year. It excludes camera, lens, lighting, carrier board, enclosure, power conversion, other system load, installation, cooling, cloud, storage, software, support and tax. It does not imply that this module can achieve five inspections per second; test the complete workload before sizing or procurement. Include optics, mounting, commissioning, storage, ongoing label review and support in a real purchase estimate.

For an annual site estimate, use camera count × (a customer quote for each complete edge computer divided by its chosen service life + measured annual computer kWh × the site's tariff), then annualize quoted one-time cameras, optics, mounts and installation over their chosen service lives; add annual recurring labor, software/support, cloud/storage/egress and tax as applicable. Use the module-only floor above only when no complete-computer quote exists; do not add it again if the quoted computer includes that module. No cost per inference is claimed without measured throughput.

Which compliance and quality limits apply?

For this narrow pilot, inspect packaging appearance rather than making medical, food-safety or seal-integrity claims. If the output will affect a regulated product's release, the responsible quality team must define and evaluate that use. Preserve lot and model-version traceability so an investigation can reconstruct the disposition. In the US, OSHA's general machine-guarding rule addresses protecting employees around machines; this visual checker is not a guard or safety control. In Canada, identify the applicable provincial or federal workplace-safety requirements for the site.

Keep worker faces and unrelated workstation activity out of the field of view where possible. If images contain people or identifying labels, document permitted uses, access and retention before upload. Canada's privacy commissioner says PIPEDA applies to commercial activities and to employee personal information in federally regulated businesses; determine whether it applies to the organization and review relevant provincial rules. In the US, review employee-video and state privacy requirements for the site. A customer-owned cloud account does not by itself establish compliance. The commissioner's employee-video case concerns an employer subject to PIPEDA, not every Canadian employer.

Before trial images are reviewed, write down the label definitions for the chosen defect family and have reviewers apply them to a shared sample. Keep threshold selection examples separate from the challenge set used to assess the final workflow. Include normal changes in print, material, lighting and alignment, plus unusable or partially obscured images. Report how many items were captured, reviewed, uncertain and escalated; a single anomaly score hides these distinct outcomes. During early operation, use an observation-only period so operators can compare system suggestions with their normal inspection process before any change to disposition authority.

Where does HyperLake fit, and when should another route be used?

HyperLake currently runs in the customer's AWS or Google Cloud account; Azure, OVH, private cloud and on-premises are coming soon. Use its published data/model capabilities for an agreed evidence and model workflow, subject to a validated integration. The edge package-inspection application remains a separate responsibility.

A standalone commercial vision appliance may be the better choice for one stable line with an established inspection task. A purely local application may be required when no package image or metadata can enter a cloud account today. Neither case needs a cloud operating platform simply to compute an anomaly score.

How should the pilot be bounded?

Run a 2-week pilot, scoped to one surface-defect question and an observation-mode acceptance plan. This is a proposed trial period, not a delivery guarantee. The team should agree who reviews uncertain cases, how long evidence is retained and who may change thresholds before a model output affects disposition.

See the physical-AI hub, the factory inspection architecture and the companion missing-component assembly review. The LLM calculator applies only to a separate language-model feature such as a reviewed summary; it does not price this vision pipeline.

Frequently asked questions

Sources

1. Anomalib PaDiM documentation
2. Basler acA1920-40gc product specifications
3. NVIDIA Jetson Orin
4. EIA Table 2.4 — average electricity price by end-use sector, 2025
5. Office of the Privacy Commissioner of Canada — employee notice example
6. OSHA 1910.212 — General requirements for all machines
7. HyperLake FAQ

Related guides

• Physical AI infrastructureWhat a Physical AI stack needs, why sovereignty matters for robot data and models, and how Hyperlake deploys and governs the stack in your environment.
• Data infrastructure for physical AI and robotics
• Robotics training-data formats and provenance