Starting with AI vision in production: start with real images

There is often no complete requirements specification yet

In practice, we increasingly see companies wanting to explore whether AI vision can add value, while they do not yet have a complete set of requirements. Usually there is a clear understanding of what is considered a good or rejected product, but not yet a complete defect definition with tolerances for each category.

Sometimes the inspection is about damage, contamination or shape deviations. Sometimes it is about mix-up detection, where the system checks whether the correct product is present in the line. In other cases, it is about recognizing product variants that operators can easily distinguish, but that have not yet been fully defined in technical terms.

The first question is simple: are images of these products already available? And are there also images of good products, rejected products and borderline cases?

If those images do not exist yet, that is where the first step begins.

Collecting images at the production line yourself

You can involve an external vision company to carry out a feasibility study. They will temporarily build a test setup, capture images and investigate whether the inspection is feasible with their software and hardware. This can be a good approach, especially when there is limited time or knowledge available internally.

One disadvantage is that such a study is often carried out within that company’s own software platform. As a result, you may already be moving towards one specific solution early in the project. Training or later retraining of the AI model may also remain an external service.

Another approach is to place a simple imaging setup at the production line yourself. This allows you to collect real production images over a period of time. These images can then be reviewed and labelled internally by quality staff or operators.

Based on that dataset, you can later test different AI vision software platforms and make a better decision about which solution is the best technical and practical fit. You are then not making a choice based on demo data, but on your own products, your own variation and your own production environment.

The advantage is that you can start quickly and learn a lot. The disadvantage is that you are already investing in hardware, without any guarantee that this camera, lens or lighting will also be the final choice. That does not have to be a problem, as long as you see the first setup as a learning setup and not as the final inspection solution.

Good AI inspection starts with good image data

A simple test setup does not have to be perfect, but the images do need to be usable. Important questions are:

• Is the defect visible in the image?

• Is there enough detail to distinguish between good and bad products?

• Does the lighting provide enough contrast?

• Does the image remain stable across different batches, shifts and product positions?

• Can quality staff label the images reliably?

If the answer to these questions is not yet clear, it makes little sense to start discussing the best AI model. First, you need to know what the camera actually sees.

Sometimes the first images show that a simple front light is sufficient. In other cases, you may find that reflections, gloss, shadows or ambient light are masking the defect. Then you need to look at a different lighting angle, another lens, higher resolution or a more controlled mechanical setup.

That is exactly the value of starting early: you discover the practical problems before you fully specify the system.

A practical test setup does not have to be complex

You do not need a powerful AI computer for the first phase. If the main goal is to collect images, a standard PC or industrial PC with enough storage is often sufficient. A powerful processor or GPU only becomes important when you want to train locally, test AI models or run real-time inspection in the production line.

To collect image material, you can start with an industrial camera, lens and suitable lighting. For simple applications, a compact camera with an integrated lens combined with a front light may already be enough to build the first dataset.

For these types of starting situations, many customers choose a uEye XC Starter Set for image-based applications or a uEye XC UVC Starter Set for video-based applications. These are 13-megapixel USB3 cameras with an integrated lens, which are easy to connect and can be used quickly for collecting image material at or near the production line.

For more critical applications, it is wise to look more carefully at the camera, lens and lighting from the start. Think of small defects, glossy products, high line speeds, low contrast or varying product positions. A short review of the field of view, working distance, resolution and lighting direction helps prevent you from collecting a large number of images that later turn out to be of limited use.

A practical way to start collecting image data

For many first tests, a simple component setup is sufficient. For example, a 13 MP uEye XC Starter Set combined with a front light and a PC or IPC with enough storage. With this, you can start collecting images for your AI vision dataset at or near the production line.

This kind of test setup is meant for learning. You capture good and rejected products, collect borderline cases and see how stable the image remains under normal production conditions. Based on this, you can later decide more effectively whether a different lens, different lighting, higher resolution or further industrial integration is needed.

View our uEye XC Starter Sets or contact us for help selecting the right camera, lens and lighting for your application.

The dataset determines the next step

Once you have collected and labelled enough images, you can better assess which AI vision solution fits your application. You can test different software platforms with the same dataset. You can see which defects are recognized correctly, where the model is uncertain and which product variations need additional examples.

This makes the choice of software, hardware and integration better founded. Not based on demos with standard sample data, but on your own products and your own production environment.

This is also an important advantage of AI vision. Training and retraining the model can become an accessible part of your own organization. New product variants, additional defect examples and borderline cases can be added to the dataset later. In this way, the system can grow along with production.

Start simple, but not blindly

Starting with AI vision yourself does not mean simply mounting a camera and expecting AI to solve the rest. The technical basis has to be right. The defect must be visible, the resolution must be sufficient and the lighting must provide the right contrast.

At the same time, you do not have to wait until everything has been fully defined in advance. By collecting images in the real production environment, you gain insight. You see where the optical challenges are, which products create discussion and which requirements will really matter later.

Our approach is practical: start with a simple but usable imaging setup, collect real production images, label them carefully and use that data to make better decisions for the next step.

Do you want to start using AI vision in your production environment? Begin with a test setup that allows you to collect reliable images. We can help you select the right camera, lens and lighting, so you can start quickly without making major mistakes in the basic setup.

Frequently asked questions about starting with AI vision

Do you need an AI computer right away to start with AI vision?

No, not if you first want to collect images. For the first phase, a standard PC or industrial PC with enough storage is often sufficient. A powerful processor or GPU only becomes important when you want to train locally, test AI models or run real-time inspection in the production line.

Can AI still recognize a defect that is difficult to see?

Usually not reliably. AI can only learn from information that is present in the image. If a defect is barely visible because of incorrect lighting, insufficient resolution or an unsuitable lens, an AI model will also struggle with it. That is why camera, lens and lighting remain important in AI vision.

Is a test setup the same as the final vision system?

No. A first test setup is mainly intended to collect images, build experience and make technical risks visible. The camera, lens or lighting may still change later. That is not a problem, as long as the first setup is treated as a learning setup and not as the final inspection solution.

Is a test setup for AI vision expensive?

No, a first test setup does not have to be expensive. Especially when the initial goal is mainly to collect images, you do not yet need a complete AI inspection system. A simple component setup with camera, lens and lighting is often enough to get started.

With a uEye XC Starter Set combined with a front light, you can already build a practical imaging setup for the first tests for under €2,000. This allows you to collect images at or near the production line, capture good and rejected products and start building your dataset.

The final inspection solution may still change later. You may eventually need a different lens, different lighting, higher resolution or industrial integration. But in the first phase, the most important thing is to start with usable images from your own production environment.

How many images do you need to start with AI vision?

That depends on the application, the number of product variants and the differences between good and bad products. For a first assessment, it is especially important to collect representative images: good products, rejected products, borderline cases and variation from normal production. Only after that can you properly determine how many additional examples are needed for training and validation.

Can you start with AI vision yourself without an external integrator?

Yes, in many cases you can start by collecting images and building a dataset yourself. However, the basic setup must be technically sound. Field of view, working distance, resolution and lighting determine whether the images will be usable later. With limited support in selecting the camera, lens and lighting, you can often start quickly and responsibly.