Model reference — Pictures
Defect detection finds the parts that are wrong on a production line — a scratch, a crack, a missing component, a misprinted label, a bad weld — and marks where on the part the problem is.
The distinctive feature of this task is that it can be learned from good examples alone. Because faults are rare and varied, a model that has been shown only correct parts can flag anything that deviates from them, which means a line can be protected without waiting to collect photographs of every possible fault. That is what makes these models practical: a few hundred images of good parts, taken under consistent lighting, and a system can be running. Consistent lighting and fixed camera geometry matter more to the result than the choice of model.
Models in this group take a single image as input: a photograph, a scan or a screenshot. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one 2-megapixel image of a part inspected for defects.
Vendor: Amazon Web Services
What it does: learns what a correct part looks like from good examples and flags any region that deviates, marking where. The most accurate anomaly method on standard industrial benchmarks and the usual first choice.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 6,000 | ≈ 24,000 | ≈ 65,000 |
Vendor: Groupe Renault
What it does: a lighter method of the same kind, quicker to set up and to run, suited to a line where inspection must keep pace with high throughput.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 12,000 | ≈ 48,000 | ≈ 140,000 |
Vendor: MVTec Software
What it does: built for speed specifically, capable of inspecting parts at millisecond latency, which is what an in-line inspection station at full production rate requires.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Vendor: Intel
What it does: a production framework carrying several anomaly methods, so alternatives can be compared on your own images and swapped without rebuilding the pipeline.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 8,000 | ≈ 32,000 | ≈ 90,000 |
Vendor: Ultralytics
What it does: a detector fitted to your specific fault types where you do have labelled examples of them. More precise than anomaly detection about what kind of fault it is, but blind to fault types it was never shown.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Vendor: Meta
What it does: outlines the exact shape of a flagged defect so its size can be measured against a tolerance, which is what turns a flag into a pass-or-fail decision.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | A100 80 GB |
| VRAM | 12 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 96 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (images/hour) | ≈ 3,000 | ≈ 12,000 | ≈ 30,000 |
What works depends on your part, your camera setup and whether you can supply examples of faults or only of good parts. Our consultants review your line and images, then recommend an approach — anomaly detection from good parts, or a fitted detector where fault examples exist — with the expected catch rate and false-alarm rate.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Industrial defect / anomaly detection. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.