AI vision brings additional opportunities

Different framework conditions, different prior knowledge

In industry, AI vision with machine learning opens up new possibilities that are not covered by rule-based image processing. © IDS

However, the fact that AI-based methods work completely differently to rule-based approaches is their greatest advantage. This enables providers to develop completely new tools that are much more intuitive to use. This means that human quality requirements can already be transferred to AI vision systems using machine learning in order to optimize and automate processes. In many cases, not a single line of source code needs to be written, making AI vision suitable for completely new target groups who no longer necessarily need to have programming skills. Feasibility analyses can therefore be carried out by the employees who have the most knowledge of products and their special features, meaning that companies are no longer necessarily dependent on programmers and image processing experts in the evaluation phase.

This means that the key skills for working with machine learning methods are no longer the same as for rule-based image processing. But here too, only sufficient specialist knowledge and experience will lead to the desired goal. Without a trained eye for the right image data, errors will also occur with this method. One of the most important factors for good results is to prepare a sufficient amount of learning data in the form of sample images for the learning process. This is where the greatest effort is required compared to rule-based methods - and also where the greatest potential for error lies.

Poor quality input leads to poor output. In order to recognize patterns and make predictions, an AI system is dependent on data from which it can learn "correct behavior". If an AI is built in laboratory conditions with data that is not representative of later applications - or worse, if the patterns in the data reflect biases - the system will adapt these biases. Data distortions, known as biases, are the result of which a neural network would make biased decisions during inference. However, with Confusion Matrix and heat maps, there are tools to make decisions and reasons for decisions visible and therefore understandable. With the help of such software tools, users can attribute the behavior and results of the AI vision more directly to weaknesses within the training data set and correct these in a targeted manner. This makes AI easier to explain and understand for everyone.

To implement a solution, it must be integrated into an application and work together with other system components, just like its rule-based counterpart. And this is often the real challenge of an application. Those who have selected the system components purely on the basis of price, performance and feature richness often have to deal with various service providers and manufacturers in terms of interfaces. Getting them to work together smoothly in a complete workflow often requires patience, time and a great deal of specialist knowledge.

With the IDS NXT complete AI vision system, IDS therefore relies on components that are optimally coordinated and can be used for all user groups with the support and experience of one contact person. This saves valuable time and expensive development resources and gets users to their individual solution faster.

Other possibilities

Increasing anomaly errors can indicate a deterioration in the condition of a system due to tool wear, dirt or other disruptive influences. © IDS

However, every application has its own specific framework conditions. It should therefore always be checked individually which task can be solved most efficiently with which tools. For example, there are many different ways to find objects. If they can be easily described and distinguished based on perspective, shape and color, a blob analysis can also deliver very good results. Using a color threshold value, segmentation then quickly provides the desired shape or position of the objects. If features can be defined so clearly and with so few specifications, there is no need for time-consuming CNN training with sample data.

If, on the other hand, the objects are more complex with more natural variance or the position of the objects varies greatly, e.g. because they are bulk goods, feature recognition is more difficult. Here, a trained network can recognize an anomaly much more effectively based on suitable sample images of different object views of good parts. Another added value of this AI method is that it can also recognize deviations from the normal case that are underrepresented in the training. In other words, those that were not planned at all. So where other methods become uncertain about something "unknown", sometimes even failing, this method is highly unlikely to miss anything. And that includes everything that may occur at some point during normal operation. Continuous data on a system status, for example in the form of increasing product faults or deviations, i.e. anomalies, enables you to determine the optimum time to maintain a system before product quality drops too much or a worst-case scenario such as a system failure occurs.

Think differently

There is no one best technology for all tasks from very different application scenarios. It is important to conduct a thorough analysis to determine which approach best fits the given circumstances. In some cases, rule-based approaches can still be effective and efficient. On the other hand, AI vision is not a trend that will be forgotten tomorrow. Its versatility to handle complex tasks with high precision for many application areas makes it extremely valuable for many companies. Its development is progressing steadily and is being driven forward by research investments and funding programs. However, it is not better and will not replace rule-based machine vision! To achieve the best results, it often makes sense to combine several approaches. If you want to use AI vision successfully, you have to be open enough to allow ideas and experiments.