In many companies, manual quality control ties up valuable personnel capacity and is prone to overlooked errors, which leads to complaints, and pseudo rejects if good parts are sorted out anyway. This article presents an approach to develop a VISION system that replaces human inspection:
In this article, we will introduce you to the steps using our sentin VISION Lab as an example. We focus on automating visual and image-based controls using Artificial Intelligence (AI) because it checks faster and more accurately than a human. In addition, an AI can detect complex error patterns more reliable than rule-based VISION systems and does not need to be constantly recalibrated. Together with our partners, we have built up a portfolio of cameras and light sources to quickly analyze the doability of applications.
The aim of such a first examination is to analyze how an image of a defect or of the object to be tested has to be taken in order to evaluate it with our algorithms. We distinguish between simple experimental setups and more complex systems that require more space and several devices. The development of such a solution is an iterative process, so these steps sometimes have to be passed through in several loops.
In this first investigation, images of objects are acquired and examined for surface and shape defects. The challenge here is to find good settings for visual recognition of the defects and to meet the requirements of the production process (e. g. cycle time, speed and environmental conditions). This includes the correct camera setting (lens selection, image resolution and exposure time), the correct selection of illumination and the estimation of the development effort required to train the AI algorithms.
Step 1: Camera Selection
We mainly work with Industrial Cameras based on the GigE Vision standard. This ensures high compatibility of the cameras and allows cable lengths of up to 100m. In this step we investigate which camera configuration is best suited for a specific application. Here we work with cameras ranging from simple 2 MP up to high resolution 20 MP cameras. For example, for easily detectable defect features or small image areas, low cost 5 MP cameras are sufficient, but for fast moving objects global shutter instead of rolling shutter sensors are required.
In addition, we also use smart cameras which, thanks to an integrated processor, are able to install and run AI models on the camera. Here, the image evaluation takes place directly on the camera hardware, which makes the installation of an industrial PC unnecessary. This system is therefore much easier and faster to integrate into the production process.
Step 2: Selecting The Lighting
Also, different lighting sources should be tested, because without the right lighting it is usually not possible to capture meaningful images. However, in some applications, such as surface inspection, the light must not cause any reflections due to the poorer detectability of defects. Therefore, the Dark Field Illumination approach can be used to detect scratches, cracks and inclusions. Structured Light (Deflectometry) is often used to detect surface and shape defects on reflective objects with complex geometry, such as painted automotive parts or plates. If a defect is present, the regular structure of the illuminated surface changes and can then be easily detected by artificial intelligence.
Step 3: Selection And Training Of The AI Algorithm
With a setup such as that of the VISION Lab, it can be checked within a very short time whether the implementation of the inspection of objects in form and illumination theory with artificial intelligence is possible. Important for the training of the AI models, however, is a sufficient database and the selection of the right AI model architecture. Often the system learns from less than 500 examples how something should look like. Depending on the application, image acquisition and evaluation can take place in a few milliseconds. For various application examples our models have achieved a detection accuracy of 99.9%. The following pictures show current studies, for example for the examination of Petri dishes and nut distribution in chocolate.
The steps presented in this article for the automatic inspection of products and components can help you to improve your quality assurance and reduce complaints and pseudo rejects. As already indicated, successful implementation may also involve going through these steps in several loops and trying out some combinations of camera, lighting and AI architecture. However, in addition to the cameras and lamps presented, there are other options and things to consider if you want to get the most out of the technology. Furthermore, these methods can be used to cover many other applications, e.g. in the packaging or metal industry.
Do you have questions about this approach? Would you like to see the sentin VISION Lab in person? If you are interested in our solutions or in digital quality assurance, you can get more information here. If you want to talk about your application, feel free to contact us.