Visual inspection AI without data scientists: the quality team's part and the hardware

What AI visual inspection needs from a small manufacturer

AI visual inspection without data scientists works when the quality team does the part it knows best: it defines the defects, supplies images of good and faulty parts under fixed light and camera position, labels or confirms them, and signs off the acceptance criteria. Training and installing the model is technical work that the station builder, a service partner or a guided tool takes on. The GPU is the smallest decision in the project.

This guide is for heads of quality, production managers and owners of small and mid-sized manufacturers: metalworking, plastics, castings, electronics, wood, textiles, food packaging. It explains in plain words what the company has to provide, how many images the published tools ask for, what to do when you have almost no defect images, and which hardware goes where. Our guide to an AI server for manufacturing covers inspection as one of three plant workloads, and our guide to a GPU server for training vision models is written for the engineers who train the models. Inspection of product quality is not a safety function; a camera system that protects people at the machine is a different case under the AI Act, as our guide to the AI Act for machine builders explains.

A few terms first. A model is the trained program that looks at an image and decides. Training means showing it many examples until it decides well. Inference means using the trained model at the line, image by image. Labelling means marking on each image what is a defect and of which kind.

Two methods: defect detection, or anomaly detection trained on good parts

There are two ways to teach a machine what a bad part looks like, and the choice depends on how many defect images you can collect.

The first is a defect detector. It learns from images in which every defect is marked with a box or an outline and a defect type, such as scratch, pore, missing clip. Ultralytics, the maker of the YOLO detection models, recommends in its tips for best training results "≥ 1500 images per class" and "≥ 10000 instances (labeled objects) per class", with images "representative of deployed environment", and about 0 to 10 per cent background images without objects "to help reduce FPs", false positives. Those figures are for best results; a first model is often trained on fewer, starting from pretrained weights, which the same page recommends for small to medium datasets.

The second is anomaly detection. It learns only what a good part looks like and flags anything that deviates. This suits small-batch production and processes with very low defect rates, where a few hundred defect images would take years to collect. MVTec AD, a public dataset for comparing such methods, shows the principle: according to MVTec, "each category comprises a set of defect-free training images", while the test set holds images with and without defects. The open-source library anomalib, under the Apache 2.0 licence, collects methods of this kind, among them PatchCore, PaDiM and, since its version 2.7.0, the few-shot method FoundAD.

Ultralytics "Tips for Best Training Results" and MVTec AD dataset page, read on 25 October 2026; the table is our summary.

Anomaly detection still needs defect images, only fewer. You need them to test the model and set its threshold, the point at which a deviation counts as a reject. Our guide to visual anomaly detection with anomalib covers the methods and their GPU needs for the IT side. Where an existing inspection machine already flags parts and too many of them are false alarms, a second model that reviews only the flagged images is often the faster start, as our guide to AOI false calls in SMT production shows for electronics.

Camera, light and part position in camera inspection

Many inspection problems start before the model: the images change while the parts do not. A shadow that moves with the time of day, a part that sits a few millimetres differently in the fixture or a lens that gathers dust each look like a defect to a model that has never seen them.

Three decisions keep the images stable. The part sits in the same position every time, in a fixture or on a conveyor with a stop. The lighting is the station's own, shielded from daylight and hall lamps, and chosen for the defect: light at a low angle makes scratches and dents visible, diffuse light suits printed or coloured surfaces, backlight shows edges and holes. The camera, lens and settings are fixed and recorded, so a replacement camera produces the same images. A machine-vision integrator chooses these components; the quality team tests them with real limit samples, parts that are just acceptable and just not, before anyone collects training images.

Labelling: who does it, and how

Labels are where the quality team's knowledge enters the model. A model learns exactly what the labels say, including their mistakes, so the people who decide on parts every day should label or check the labels, against a written defect catalogue with photos of limit samples.

Ultralytics states the rule plainly: "All instances of all classes in all images must be labeled. Partial labeling will not work." In practice that means one person labels, a second checks a sample, and where the two disagree the defect catalogue is clarified along with the image. For anomaly detection the task is lighter: someone confirms that every training image shows a good part, because a defective part among the "good" images teaches the model that the defect is normal.

The labelling tool can run on your own server. CVAT, an open-source annotation tool whose core is under the MIT licence, runs on your own infrastructure with Docker Compose and supports boxes, polygons and masks; its maker lists built-in automatic labelling with SAM 2 and SAM 3 for its paid online and enterprise plans. Images of your products and their defects then stay in the company.

From images to an approved model

The station builder or service partner trains the model on the server and hands back three things: the model, its results on images it has not seen during training, and the threshold proposed for the line. The quality team then decides with figures, not impressions.

1. Keep a test set apart. Before training, put aside a share of the images, with as many real defects as you have, that the training never sees.
2. Count both errors. A false reject sends a good part to scrap or rework; an escape lets a faulty part through. The threshold trades one against the other, and only quality can say which costs more for each defect type.
3. Sign off acceptance criteria. For example, every escape on the test set reviewed, the false reject rate below the current manual rate, a defined run at the line in parallel with the manual check.
4. Version the model per product. Each product variant, camera and model version is recorded, so a complaint can be traced back to the model that judged the part.
5. Plan retraining. A new product, a new supplier of raw material or a changed surface finish means new images, new labels and a new acceptance run. Collect rejected and borderline images at the line from the first day.

The training is the shortest part. On a server with RTX PRO 6000 Server Edition cards a team can try several model sizes or settings overnight, one per card, as our training server guide explains; we found no published training time for an inspection model on these cards, so a test with your own images gives the figure. We supply the training server and the line-side GPUs; send us your parts, defect types, cycle time and the images you already have, and our team replies with a configuration for training and for the line.

Who does what in a project without data scientists

Our recommended division of tasks; "integrator" means the station builder or service partner, and your contracts define the actual split.

The hardware: at the line and in the server room

Inspection uses hardware in two places. At the line, a small GPU runs the model on every image at line speed and keeps running when the network to the server room is down. In the server room, a larger server trains and retrains the models and keeps the image archive.

NVIDIA L4 product page (24 GB, 300 GB/s, 72 W, single-slot low profile, PCIe Gen4) read on 25 October 2026; RTX PRO 4000 SFF and RTX PRO 5000 from NVIDIA's product pages as listed in our GPU guides; server power by our estimate (cards, processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

The L4 is passively cooled and needs the airflow of a server, as NVIDIA's product brief states and our comparison of the RTX PRO 4000 SFF, RTX PRO 2000 and L4 explains; in an industrial PC or a cabinet without such airflow, the actively cooled RTX PRO 4000 SFF fits where the PC's maker lists it. Which card a station needs depends on your model's time per image at your cycle time, so the station builder measures it on the card before you order for every line. Serving a trained model efficiently at the line is a topic of its own, covered in our guide to serving vision models with Triton, ONNX and TensorRT.

The training server is the entry configuration of our AI servers: two NVIDIA RTX PRO 6000 Blackwell Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. Each card splits into up to four MIG instances of 24 GB, so several small training jobs can share one card, as our RTX PRO 6000 MIG profiles show. Inspection training runs now and then, after a product change or when new images arrive, so the same server usually also carries the company assistant or other AI work for the rest of the time.

Licences to check before the first model

The software licence matters more than people expect in inspection, because the trained model runs inside your production. Ultralytics' licence page states that "All Ultralytics YOLO trained models fall under the AGPL-3.0 License by default" and lists its Enterprise licence for "Internal business tools or private company applications" and embedded deployments in "edge devices" and "cameras". Our guide to the Ultralytics YOLO licence sets out what that means for an in-house station. anomalib and the CVAT core are under the permissive Apache 2.0 and MIT licences. Datasets have licences too: MVTec states that its dataset may not be used "for commercial purposes". Your legal counsel confirms which licences fit your use; the GPU plan is the same whichever you choose.

What we supply

We supply the NVIDIA L4, RTX PRO 4000 SFF and RTX PRO 5000 for line-side and lab systems, after a check that the card suits the chassis, power and cooling, and build AI servers to order for training, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component handled through us. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the platform on the training server and the MLOps around it; the camera station itself comes from your machine-vision integrator. You can buy this hardware from us: our team sizes the configuration to your workload and budget and replies within one business day. Describe your inspection task and the stations you plan, and we answer with the hardware for each place.

FAQ