What does a computer vision consultant do for visual inspection?
An inspection model is useful when it makes the same call your best inspector would, on the parts your line actually produces. So I start with the decision instead of the model: which defects matter, what counts as a reject, and how fast the answer has to come back. Then I review your images and labels, train a first model, and measure it on images it has never seen. The result is counted as missed defects and false rejects at the threshold you will run.
The work runs all the way to production. I train object detection, segmentation and anomaly detection models in PyTorch or TensorFlow, with YOLO, RF-DETR and OpenCV where they fit, and deploy them with Docker and Kubernetes on cloud, on-prem or edge GPUs. I've spent a lot of time on the unglamorous part, where a model that behaved perfectly on a laptop has to survive a real line.
| Stage | What we check or deliver |
|---|---|
| Scope | Defect classes, the reject rule, the cycle time and the target hardware |
| Data review | A sample of good and defective images, label quality, and a held-out test set |
| Baseline | A first model and its misses and false rejects on that test set |
| Training | Documented training runs, with dataset versions and configuration |
| Evaluation | Decisions at your operating threshold, plus a review of every failure |
| Deployment | The model packaged and timed on the hardware it will run on |
Which computer vision tasks fit defect detection?
Most inspection problems are one of four tasks, and the choice decides what you need to label.
| Task | What it answers | Output |
|---|---|---|
| Object detection | Where is the defect, and which kind is it? | A box and a class per defect |
| Segmentation | Exactly which pixels are defective, so size or area can be measured | A mask per defect |
| Anomaly detection | Does this part look different from good parts? | A score per image or region |
| OCR | What does the label, code or marking on the part say? | Text |
Object tracking and real-time video cover parts that move past a camera instead of stopping under it.
Can I see real AI visual inspection results?
The client results I can state are short, because the images, models and customers stay private. Automated visual defect detection cut inspection time by roughly 80% for a manufacturer, and I took a production inspection model from about 70% accuracy to 99.999%. Treat both as proof that a jump of that size is possible, not as a forecast for your line. They depended on those parts, defects and labels.
The part you can inspect in full is a measurement. I maintain a training and serving engine for RF-DETR defect-detection models, and I published how its faster variants compare: three RF-DETR models fine-tuned on an industrial defect-detection set, timed on an NVIDIA RTX 5090 with TensorRT 10.12 on 12 October 2026. TensorRT FP16 ran the network about 5 times faster than PyTorch FP32 and moved mAP50:95 by 0.002 or less. The post also shows the variants that did not pay off, including an INT8 engine in which TensorRT ran every layer in FP16, and the method I use so a comparison measures the model and nothing else.
How accurate can AI visual inspection get?
There is no general figure. It depends on how visible each defect is in your images and how consistently they are labelled, and only a held-out test from your own line can show it. If a target looks unrealistic for your images, I'll tell you early.
I report results as decisions at your operating threshold, alongside mAP (mean average precision, the usual detection score). In inspection, a changed decision is a missed defect or a false reject, and a small mAP change can hide detections that flip between found and missed. In my RF-DETR tests, one CPU quantization method moved mAP50:95 by at most 0.014, yet changed 1 to 4 detection decisions per model on a 10-image test set. A test set that small gives a direction, not a precise figure, so a real evaluation needs enough defective parts to count both kinds of error.
What do I need to send to scope a visual inspection project?
Send a representative set of images of good and defective parts, any labels you already have, the list of defect types and the rule an inspector uses to reject a part. Add the cycle time per part, the camera, the hardware you plan to run on, and whether the images may be used for training. If a model is already in place, its misses and false rejects are the most useful thing you can send.
There is no useful fixed price or schedule before that review. Labelling and evaluation can take more work than the training run, so the quote covers data work, testing and deployment as well as GPU time.
How fast will the inspection model run on the line?
Your cycle time sets the target, and the model is often not the slow part. In the RF-DETR measurements the test frames were 16 megapixels, and the TensorRT network took between 1% and 5% of each request on the GPU host. Preprocessing the frames took most of each request. So I profile the whole request before optimizing the network. When the model itself has to get faster, that is inference optimization work with TensorRT, ONNX and Triton.
Frequently asked questions
How much does an AI visual inspection project cost?
I quote after seeing sample images, the defect definitions, the accuracy target and the hardware. Labelling and evaluation can take more work than the training run, so a description of the part alone is not enough for a useful estimate.
How many images do I need to train a defect detection model?
There is no fixed number. It depends on how many defect types you have, how visible they are and how much the parts vary. I review a sample first and keep a separate test set with enough defective parts to count misses and false rejects, because a very small test set only shows a direction.
What accuracy should I expect from AI defect detection?
Only a test on held-out images from your line can show it. One production inspection model I worked on went from about 70% to 99.999% accuracy, but that result depended on its parts and labels. I report accuracy as missed defects and false rejects at the threshold you will run.
Can the inspection model run on our existing hardware?
Often, and I time it there before handoff. I have measured RF-DETR defect models on an NVIDIA RTX 5090, where TensorRT FP16 ran the network about 5 times faster than PyTorch FP32, and on a CPU without VNNI, where PyTorch dynamic INT8 sped the network up 1.3 to 2 times. I haven't published Jetson measurements, so on one I would measure first.
How long does a computer vision project take?
Training is only one part of the schedule. I need to review the images and labels, set aside a test set, train a baseline, inspect its failures and time the model on the target hardware. I can estimate the timeline after that first data review.
How does a visual inspection engagement start?
With an intro call about the part, the defects and the hardware, followed by a sample of images. From those I can tell you whether the accuracy you need looks realistic and what the first model should be.