10 Questions about AI Vision

Machine learning in particular opens up completely new possibilities that cannot be covered by conventional rule-based image processing. Nevertheless, it is crucial to evaluate the technology realistically and critically in order to clearly understand its actual applications and limitations. Because it's obvious that AI is not a temporary trend that will soon be forgotten. Its ability to handle complex tasks with high precision in many application areas makes it extremely valuable for companies. Its development is constantly advancing and is being further accelerated by research investments and funding programmes. Looking forward, AI will be indispensable as a game changer in almost all areas of life.

How does AI Vision work in industrial applications?

Especially in industrial applications, AI-based image processing relies mainly on machine learning methods. In this approach, computer programmes are empowered to learn from experience and automatically extract patterns and insights from data without explicit programming. This is done by adapting algorithms and models on data to enable prediction, pattern recognition and decision-making.

AI-based methods demonstrate their power especially in image data with highly variant content. Patterns and features are recognised that could hardly be clearly defined as a recurring shape, colour or position with rule-based image processing. However, the object features relevant for recognition are no longer specified by a predefined programme sequence. Neural networks are taught in a learning phase to associate them with labels through repeated seeing. This often requires a large number of example images of the learning content. The more variant these are, the more stable the machine learning algorithms are in recognising their relevant features in regular operation, even in unfamiliar scenes.

What is the difference to previous methods or approaches?

Rule-based image processing algorithms are often developed specifically for one particular application and are difficult to migrate to new tasks. AI models, on the other hand, can be trained on one task through "transfer learning" and then transferred to related tasks without being retrained from scratch. This facilitates the reuse of models and speeds up the development of new machine vision applications.

However, the key skills for working with machine learning methods are no longer the same as for rule-based image processing. The decisive factor for the results' quality is no longer the product of a manually developed program code by an image processing expert, but is determined by the learning process with suitable sample data. This requires a deep understanding of the application being used. With the right tools, feasibility studies can then be done by domain experts alone, who themselves have the most knowledge of products and their specifics. Companies are thus less dependent on programmers and image processing experts in the evaluation phase.

Which industries will benefit from AI Vision?

AI systems can classify images into different categories, which is very helpful in applications such as image recognition or the identification and classification of products. In addition, AI-based image processing can automate many tasks that were previously performed manually by humans, such as identifying defects or sorting objects on conveyor belts. Specifically, the ability to identify complex patterns and structures in images, even when they are difficult to see by the human eye, makes it an important tool in quality assurance. Overall, the integration of AI-based image processing in these industries leads to an improvement in efficiency, quality, safety and cost-effectiveness.

Can AI Vision support quality assurance?

Specifically, anomaly detection methods can be trained very effectively to identify defective products or components by detecting irregularities, cracks, deviations or other defects in images. This enables instant detection of defects and sorting out of defective products. Early detection of quality problems and instant sorting out of defective products reduces waste, which saves costs and increases productivity. As the anomaly method detects both known and unknown deviations, such as wear patterns, AI Vision can also be ideally used to predict maintenance needs in machinery and equipment. Any indication of potential problems contributes to predictive maintenance, minimising unplanned downtime.

How does AI affect embedded vision?

The fact that AI-based methods work in a completely different way enables manufacturers like IDS to develop new and intuitive development tools for image processing. With them, human quality requirements can already be transferred to image processing systems today. Especially very complex embedded system development, which has required special expertise up to now, benefits from this. The IDS NXT AI camera system is a good example of how easily vision processes can be developed and put into operation on a small PC-independent system. The very fact that a large part of the development and evaluation process can be done in an easy and intuitive cloud service, without the need for specialised experience in AI, application programming or image processing, brings embedded vision closer to whole new audiences.

Furthermore, AI-based algorithms can be parallelised very well, i.e. effectively accelerated with appropriate hardware to process large amounts of data in real time, and not only with powerful GPUs (graphic processing units) in large data centres. With emergence special NPUs (neural processing unit), AI vision can also be executed very energy-efficiently with small embedded vision devices. This enables a scalable use of the technology, depending on the application requirements on different hardware platforms.

Will AI Vision contribute to the industry's sustainability goals?

With the ability to monitor and target production processes, companies can use resources such as water, raw materials and energy much more efficiently. This helps to reduce waste and scrap, which in turn saves resources and energy. By improving the efficiency, quality and sustainability of production processes, AI-powered industrial cameras can help minimise the industry's environmental impact while increasing economic profitability.

What are the limits of AI Vision?

It is hard to say what are the limits of a technology as long as it is still developing so strongly and lacks experience. The limits of AI-based image processing become visible, for example, when the desired results are not achieved. This is not necessarily due to technological reasons, but is usually caused by a lack of experience with AI methods. The greatest effort and at the same time the greatest potential for error compared to rule-based methods lies in providing sufficiently good and appropriate sample data for the learning process. Poor quality input leads to poor output. An AI system depends on data from which it can learn "correct behaviour". If an AI is built under laboratory conditions with data that is not representative of subsequent applications, or worse, if the patterns in the data reflect biases, the system will adapt these biases and make biased decisions during inference.

However, even if neural networks can be reliably trained for many visual tasks, limitations and challenges always arise. It is therefore important to have realistic expectations of the capabilities of AI systems and to recognise that in some cases they can complement, but not completely replace, human expertise and interpretation.

What challenges arise from industrial AI projects?

It is therefore important to realise that AI is neither magic nor so intelligent that it can anticipate what we expect from it. AI-based image processing is a powerful tool if it is used correctly. To do this, it is once again important to clarify exactly what the task of a machine vision system should be. The more clearly the question for a certain result is stated, the more precisely the appropriate learning content can be prepared for the training process. The challenge further lies in attributing undesired results to the learning process, i.e. to the transfer of knowledge, instead of trying to precisely control the decision-making process, as was necessary in the rule-based system. Inference, which many call a black box, is just the sum of the right input data. The better you train the system, the more likely the expected outcome. Some new thinking is necessary for this.

The way a trained neural network works is based only on statistics, probabilities, i.e. mathematics - but it is difficult to understand purely by human reason. With Confusion matrix and heatmaps, however, there are tools to make decisions and reasons for decisions visible and thus understandable. With the help of such software tools, users can trace the behaviour and the results of the inference more directly back to weaknesses within the training data set and correct them systematically. This makes AI more explainable and comprehensible for everyone.

How can AI Vision evolve in industrial automation?

There is no single best technology that is suitable for all application tasks. It is important to perform a detailed analysis to determine which approach is best suited to the given circumstances. In some cases, rule-based approaches can still be effective and efficient. On the other hand, machine learning's ability to handle complex tasks with high precision is extremely valuable to many organisations. However, it is not better in all respects, nor will it replace rule-based image processing! To achieve the best results, it therefore often makes sense to combine several approaches. Those who want to use AI successfully should be eager to experiment and open for new ideas and approaches. Thinking differently will pay off in a "return on investment" in the long run.