EN
Let me start with some honesty. Adding a new product line is a nice milestone, but it can also bring a nasty headache. Your vision inspection system does not know the new parts. It also has no idea what counts as a good part or a bad part. You have to teach it. If you do this the wrong way, you will end up with false rejects, missed defects, and a lot of frustrated operators. The good news is that training a vision inspection system can be quick and painless if you follow the right steps.
Preparation Starts with Sample Collection
Before you even touch the system, you need to gather your samples. This is the most important step, and yet it is the one that people rush through the most. Do not be that person.
You need multiple types of samples. First, collect good parts. You will need more than just a few images of products that meet your quality standards. You might need dozens or even hundreds. The more images you have, the better the system will understand what acceptable looks like. These good parts should also cover the normal variations you expect to see, such as slight differences in color, small shifts in position, or minor texture changes. If your training only includes perfect parts, the vision inspection system will reject everything that looks slightly different, even if it is still perfectly fine.
Second, collect bad parts. Make a comprehensive list for each defect type that matters to you. This list could include scratches, cracks, missing components, wrong colors, or misaligned labels. The wider the variety of defect examples you have, the better the system will learn what to look for. One study showed that a pre trained defect detection model could adapt to new products after receiving a moderate number of new samples and a small amount of fine tuning. But moderate does not mean minimal. A poor dataset will still yield poor results, so make sure you prepare well defined samples.
Use Quick Teach for a Fast Start
Once you have your sample dataset ready, it is time to teach the system. You are in luck because almost all modern vision inspection systems have a built in feature called Quick Teach or one button teach. This makes your job much easier when adding a new product line.
Quick Teach works by taking a reference image of a defect free part and automatically setting the system parameters based on that image. The system looks at the reference, measures the key features, and applies tolerances. You do not have to manually enter numbers or guess where the thresholds should be. The system does the heavy lifting for you.
This method works well for quickly inspecting a batch of parts that all look the same. For example, if you are inspecting identical stamped metal parts or molded plastic components, Quick Teach can get the job done in a few minutes instead of a few hours. Some modern cameras reduce setup time from hours down to just minutes by automatically learning from a few sample images. The key is to make sure your reference image is a good representation of the appearance of the parts you expect to see in the batch.
Move to Standard Teach When Parts Vary
Quick Teach has a drawback. If your parts show significant variability in appearance, Quick Teach will tend to reject a larger number of good parts. In that case, the system needs to learn from a number of iterations of parts.
Standard Teach works differently. Instead of using a single reference image, you run a whole batch of good parts through the system. The vision inspection system looks at all of them, measures the natural variation, and sets its tolerances to include that whole range of acceptable results. This way, the system learns what good looks like across your actual production conditions, not just in a perfect lab setting.
There is no limit to how many samples you can use during the Teach process. The more good parts you show the system, the better it understands what is acceptable. And here is a pro tip. Only use good parts when you do a standard Teach. If you accidentally include a bad part, the system will learn that the defect is acceptable, and you will have a mess on your hands.
Manually Adjust When Needed
Sometimes automatic teaching is not enough. Your parts may have complex features that the system struggles with. In those cases, you need to manually adjust the parameters.
Most vision inspection systems allow you to go into the tools and adjust the settings by hand. You can adjust the region of interest, change the tolerance thresholds, and fine tune the pass or fail logic. This takes more time and requires some experience, but it gives you the most control. For high precision applications where even tiny errors matter, manual adjustment is worth the extra effort.
If you are working with a new product that is similar to an old one, you can sometimes save a lot of time by using an existing inspection as a starting point. A study on transfer learning showed that a model trained on older products could be adapted to new ones with only moderate amounts of new training data, maintaining over 98 percent accuracy. That means you do not always have to start from scratch. You can build from existing models instead.
Leverage AI for Complex Inspections
Traditional vision inspection systems rely on fixed rules. Is the scratch longer than X millimeters? Is the color outside the range of Y? That works fine for simple inspections. But when defects are irregular, inconsistent, or hard to describe with fixed rules, conventional systems tend to struggle.
AI powered vision inspection changes the game. Instead of following a prescriptive set of rules, the system learns from examples. You show it hundreds or thousands of good and bad images, and it figures out the patterns on its own. AI powered vision inspection is especially beneficial for manufacturers trying to preserve product quality and minimize waste. It is also highly adaptive to product changes. Some platforms combine supervised and unsupervised AI learning with rule based tools to reduce false positives and prevent defective products from reaching customers.
The great thing about today's systems is that they are user friendly. Certain cameras now include built in AI that automatically learns from just a few sample images, eliminating the need for lengthy manual configuration. User friendly interfaces guide operators through every stage of the setup process, from adjusting lighting to registering reference images. Even operators with limited vision inspection experience can create robust inspection processes.
Testing and Validation Are Essential
This is where most people make critical mistakes. They train the system, run a quick test, and then put it straight into production. That is a huge mistake.
Before you let your vision inspection system run unsupervised, you need to validate it properly. Run a batch of known good and known bad parts through the system. See if it catches all the defects. See if it rejects any good parts. If your false reject rate is too high, go back and adjust your tolerances. If your miss rate is too high, add more defect samples to your training set.
A proof of concept approach works well here. Start with one critical inspection point on your new line instead of trying to automate everything at once. Gather your golden dataset of good and bad samples, test the feasibility, and only then scale up. Doing this step correctly will ultimately save you a lot of money.
Plan for Continuous Improvement
Training a vision inspection system is not a one time event. Your products will change. Your suppliers will change. Your production conditions will change. Your inspection system needs to keep up.
Build a process for ongoing training. Whenever you find a new type of defect that your system missed, add those images to your training set and retrain. Whenever you change a product design, update your reference images. Some advanced systems offer continuous learning capabilities that adapt to product variations over time. The more you train your system, the smarter it gets.
Manufacturers who get this right see real results. AI vision inspection enables earlier defect detection, faster root cause analysis, and real time actionable insights that boost efficiency and reduce waste. YIHUI designs vision inspection equipment with these training principles in mind, helping manufacturers across machinery, electronics, aerospace, and automotive industries get new product lines up and running faster. After all, a vision inspection system is only as good as the training you give it. Do it right, and it will protect your brand for years to come.