The limits of artificial intelligence

Why the use of AI is not sufficient for image processing applications

AI-supported image processing has proven to be a good tool for many applications to detect errors that are easily visible to the human eye, but difficult to describe with a set of rules.

Until now, such tasks could not be solved, or only with tremendous effort. Nowadays AI technology also makes it possible to solve some tasks faster and easier than before using rule-based image processing.

The use of artificial intelligence (AI) in machine vision

AI-based solution to detect ripeness of tomatoes

Artificial intelligence is a field of research that aims to develop machines that can learn, plan, and solve problems using intelligent algorithms.

An AI-based system is trained using a large number of examples and then tries to generalise what it has learned to solve new tasks, in the case of machine vision applications for example the entire variance of a production.

The ability of AI to process large amounts of data quickly and learn from previous experience also makes it a valuable tool for machine vision. In practice, an AI system is fed with lots of images that contain both objects and their surroundings. The system is then able to identify the objects it is looking for and ignore the surroundings. An AI system trained in this way can then be applied to new images and identify the desired objects with high accuracy.

AI-based detection of different kinds of bakery

Applications and AI algorithms used in machine vision

Artificial intelligence can analyse images for their content to find defects, extract edges or classify objects and is mainly used in applications where the object is highly variable and/or has non-specific features or defect types.

Classification

Based on previously defined characteristics, objects are assigned to a class

Object detection

Trained objects are located and identified, framed with a rectangle and labelled

Semantic segmentation

A class is assigned to each pixel in an image. Trained defect calsses can be localised with pixel accuracy

Anomaly detection

Images of defect-free objects are trained; all deviations are detected

Edge extraction

Enables the extraction of edges that cannot be identied with conventional filters

Instance segmentation

Combines object detection and semantic segmentation and is particularly useful when objects are close together, touching or overlapping

Deep OCR

Robust localisation of characters regardless of orientation, font, and polarity

Pre-trained networks (CNNs = Convolutional Neural Networks and RNNs = Recurrent Neural Networks) make it possible to develop applications even with a relatively small number of training images, the algorithm analyses the trained images and automatically learns which features can be used for identification.

Limitations of AI for machine vision applications

AI accuracy is affected by factors such as the quality of the image, the amount of information contained in the image, the image resolution and the illumination. In practice, the AI is trained to a specific image resolution and can therefore only be as good as the images it is trained on. If the image quality is poor and the resolution is low, it is difficult for the AI to accurately analyse the content of the image. This may be due to a lack of detail in the image, or visual noise, such as dark or light spots on the image, which can make object recognition difficult.

In machine vison applications, however, it is important to start exactly at these "spots" on an image in order to determine which of them are errors and which one are not. This is where complex training processes come into play, as these "spots" have to be framed, assigned to a class and it has to be ensured that no area in the image is overlooked or not marked. If, for example, two features are still mixed up after this training process, further images with exactly these features must be used for training and it is extremely important to assign the respective class correctly. Sometimes it is also necessary to delete pictures from the database that lead to confusion. However, this training, necessary optimisation and resulting decisions require a lot of experience and a high personnel effort.

Poor lighting conditions also often cause problems for an AI-based approach and the analysis of poorly lit images is difficult. Even with defects that are not very clearly visible, it is only possible to distinguish between production variance and actual defects after extensive training.

In our market, however, there is no other option than to proceed exactly in this way. An exception is the use of so-called anomaly detection, a method that does not recognise the error type, but only that a product variance is present. This may or may not be a defect.

This is exactly where humans come into play again. If the detected error is a real defect, no further "manual input" is needed. In all other cases, however, retraining is necessary. This procedure requires a lot of experience and care and usually makes the use of AI relatively expensive. The advantage, however, is that no special programming knowledge is required, which significantly increases the circle of users.

In combination with classical image processing algorithms and a skilful application of trained data, however, semi-automatic training can be generated very well, which saves time and money.

Advantages of combining AI and human expertise

By combining artificial intelligence and human expertise, images can be analysed more accurately and quicker than it would be the case with manual analysis. This can help companies analyse all images, even those with poor resolution or quality, to make better-informed decisions. AI can also be used well to quickly analyse a large number of images and save the user a lot of time.

illustration showing the combination of AI and human expertise

AI technology is certainly a powerful tool, but it is not sufficient on its own to guarantee reliable image processing. As AI cannot measure, it will still be necessary to combine AI technology with classical technology. In addition, AI technology is only as good as the data available. It also requires a sufficient amount of image data for training, which must be properly labelled.