Distinguishing bubbles from particles using artificial intelligence (AI)
A major challenge in the automatic visual inspection of parenteralia is the differentiation between air bubbles and particles. We solve this with the support of artificial intelligence. This significantly reduces the false reject rate.
Many influences can affect the detection performance of a machine for automatic visual inspection (AVI). One of the biggest challenges is the generation of bubbles in the filler. Although these bubbles do not pose a risk to the patient, they do present AVI with a major challenge: distinguishing these bubbles from critical particles. High machine speeds, handling, the filling process and the potential tendency of the filler to bubble are common challenges.
As a result of the difficulty in distinguishing between bubbles (non-critical) and particles (critical), AVI will produce an excessive number of rejects or retests. In both cases, this represents a significant business risk. Either time-consuming manual retests have to be organized, or in case of a wrong bad rejection, many good product will be unnecessarily rejected.
In cases where it is difficult to distinguish between bubbles and particles, AI can significantly exceed the performance of a classical image processing algorithm.
In a large-scale pre-engineering study with an industrial partner we were able to demonstrate this. A large batch of samples was provided by the customer. Artificial defects were inserted, similar to a Knapp test set. Through a continuous serialization process these could then be traced back by the customer and the classifications could be compared 1:1. At the customer's site a series test with manual inspection was carried out, at our site a series test was performed on a high performance machine with 600 / min throughput. First of all, the algorithms of classical image processing were adapted to the measuring task. These could map the performance of the MVI, except for the differentiation of particles and bubbles. However, by means of a subsequently trained, supervised AI, the detection performance was dramatically improved to that of the classical algorithm.
Therefore, our approach is a profound preliminary study in which, in addition to the design of the optical components of the image, the image processing is already preconfigured. Part of the preliminary study is then also the determination of the detection rates that can be achieved later. This includes the classical (image processing algorithms) as well as the use of AI. Many tasks can be solved with sufficient quality using classical image processing algorithms, while others only become relevant for production use through the use of AI due to their severity.
- Vision system design process
- Latest vision components (hard and software, including AI)
- Task specific vision setup
- Reduction of false rejects
- Increase of production outcome
- Increased profits