The first question in almost every AI inspection project is how many defect images the model needs, and the honest answer is that it depends on the method. A supervised model needs labelled examples of every defect class. An anomaly model can start from good parts alone. On a line that makes one bad part in two thousand, that difference decides whether you start next week or next quarter. This guide covers image counts, labeling, rare defects and keeping models accurate once they are live. Size your own dataset in a short model training demo.
Start With Good Parts, Add Defects as They Arrive
iFactory Model Lifecycle gives your team labeling tools, anomaly models for rare defects and drift monitoring, with retraining run on the on-site NVIDIA server. Inspection starts on good images while the defect library grows.
Two Ways to Train, and What Each Asks of You
Supervised models learn from labelled examples of each defect, so they can name the defect and often show where it is. Anomaly detection learns what normal looks like and flags anything that differs. PatchCore, presented at CVPR 2022, is a well-known example: it fits a memory bank of patch features using defect-free images only. Many lines run both, an anomaly model from day one and supervised classes as real defects build up. Talk through the right mix with our vision data team.
- Needs labelled images of every defect class
- Names the defect, so it can grade and route it
- Struggles with defects it has never seen
- Best for frequent, well-defined defects
- Fits on good images, which every line has plenty of
- Flags the unusual, including new defect types
- Says something is wrong, not what it is
- Still needs real defect images to test it
What Vendors Say About Image Counts
Published numbers vary widely, and most are minimums to start training rather than what a production model needs. Treat the rows below as vendor and research guidance, not rules. The counts that matter are the ones your own validation results support. Check your defect list against them in a dataset sizing call.
Worked Example: What a 1-in-2,000 Defect Rate Allows
A good line is a slow source of defect images. The example uses illustrative figures for a stamping line to show how quickly one defect class builds up. The same week produces tens of thousands of good images, which is why an anomaly model can start long before a supervised one. Run the sums for your own line with our support engineers.
Classification, Box or Mask: Use the Cheapest Label That Works
The label type sets both the labeling effort and what the model can tell you. A whole-image label says a defect is present. A bounding box says where. A segmentation mask outlines the defect pixel by pixel, so its size can be measured against a spec limit. Choose by the decision the line has to make, not by what looks most advanced. See all three label tools in a labeling demo.
Know When Each Defect Class Is Ready to Train
Model Lifecycle counts labelled images per class, checks label agreement and tells you when a class has enough real examples. Until then, the anomaly model keeps watching every part.
A Label Guide That Keeps Two Labelers in Agreement
Inconsistent labels teach the model to be inconsistent. Write one guide per defect class before labeling starts, then have two people label the same sample and measure agreement. Under AIAG guidance, a kappa of 0.75 or more counts as good agreement, and 0.90 is better still. Low agreement usually means the definition is unclear, not that the labelers are careless. Ask our support team for a blank guide.
Augmentation and Synthetic Defects Help, Within Limits
When real defects are scarce, there are two ways to stretch them. Augmentation varies existing images. Synthetic data renders or pastes new defects onto images of good parts. NVIDIA's own worked example, training a scratch model on rendered images, showed the usual catch: long scratches were missed and reflective edges were flagged until real, hand-labelled images were added. Plan the right blend in a data strategy session.
Split Train, Validation and Test by Lot and Day
The training set fits the model, the validation set tunes thresholds and settings, and the test set is looked at once to give an honest score. A random split by image leaks: neighbouring frames and the same part shot twice end up on both sides, and the score mostly measures memory. Split on the unit that will be new in production, such as lot, shift or day. Get a split check from our vision data team.
Six Signals That Should Trigger Retraining
A model that never changes still meets a line that does. New material lots, a replaced light, die wear and new defect types all shift the images it sees, which is called data drift. Model Lifecycle watches the score distribution on known-good parts and the share of parts sent to review, then retrains on the on-site NVIDIA server and compares the result on the frozen test set before release. See drift checks in a live walkthrough.
Price the Labeling Before You Start
Labeling time is the cost most projects underestimate, and it depends mostly on label type. The worksheet uses assumed times per image to show the arithmetic only, so time a batch of your own images before planning. Our support team can help you scope it.
Multiply by your own labour rate. If whole-image labels answer the line's question, the same dataset costs a fraction of the mask effort.
Labeling, Anomaly Models and Retraining on One NVIDIA Server
iFactory delivers Model Lifecycle with the hardware it runs on. A pre-configured NVIDIA AI server ships racked and ready with labeling tools, anomaly models and drift monitoring pre-loaded, so images, labels and retraining stay inside the plant. You rack it, plug in power and Ethernet, and the AI is live. Scope it for your lines in a deployment scoping call.
- Pre-configured NVIDIA AI server, racked and ready
- Labeling tools with double-label agreement checks
- Anomaly models for rare defects from day one
- Supervised classes as real defects build up
- Test sets split by lot and frozen
- Drift monitoring and on-site retraining
- 24×7 remote monitoring of models and hardware
Cameras and server in, good images flowing, a label guide written for each class.
Anomaly model in shadow, first labels checked for agreement, test sets frozen.
Models released, retraining triggers set, QA trained on labeling and review.
Model Lifecycle Tools Built for the Plant, Not the Lab
Model Lifecycle puts dataset work in the hands of the quality team that knows the defects. Our support team can walk you through a pilot plan.
What Plant Managers Ask About Training Vision Models
Train on the Images You Have, Grow Into the Ones You Need
Start inspecting with good-part models, label defects as they arrive and retrain on site when the line changes. Delivered turnkey and live in 6 to 12 weeks.
