The photo shows machine vision systems that automatically control work performed by people or other machines.

Artificial intelligence in vision systems.

Artificial Intelligence (AI) is a technology that changes every area of ​​life. It is a comprehensive tool that allows people to rethink the way we integrate information, analyze data, and use the insights thus obtained to improve decision making.

Deep Learning (DL) techniques are a sub-field of machine learning within the scope of AI, and their operation is often compared to learning the human brain, but these terms are mathematics. In recent years, a large increase in the use of deep learning techniques is estimated. Vision systems, as they offer many possibilities in heavy-duty applications to be realized with traditional vision system tools.

Intelligent systems vs. classic algorithms.

Deep Learning has pushed the boundaries of what was possible in the field of digital image processing. However, this does not mean that after the advent of Deep Learning, the traditional machine vision techniques that had developed in the preceding years have become obsolete.

Traditional techniques can solve the problem much more efficiently and with fewer lines of code than DL. Rule-based algorithms need a known set of variables for a given inspection: which elements and in what variants do they occur? what are the dimensional tolerances? does the inspection object change its location?

When we are dealing with a controlled environment and all the relevant variables are known, traditional algorithms will be able to fulfill their role not only precisely but also efficiently, performing up to thousands of inspections per minute.

Machine learning uses algorithms to discover patterns in data sets, which the program then adjusts to perform. An important part of machine learning is the set of elements on the basis of which the algorithm will be able to learn and then make decisions. The aforementioned set usually consists of hundreds of examples, depending on the purpose of the system, which translates into the accuracy and resistance of the algorithm to changes. Deep Learning may offer greater accuracy and versatility, but it requires significantly more resources and computing power, which translates into processing time and costs.

Advantages of Deep Learning.

Compared to traditional machine vision techniques, DL enables greater accuracy in tasks such as image classification, semantic segmentation, object detection, and simultaneous localization and mapping. Because the neural networks used in DL are learned rather than programmed, applications using this approach often require less specialized analysis and tuning. DL also provides excellent flexibility as neural network models and structures can be retrained using a custom dataset for any given case, unlike CV algorithms which are usually rigid for any given case.

Systems based on artificial intelligence are good for problems related to the inspection of surfaces, food, difficult OCR cases, or others, when the product has high variability and / or its flaws are very diverse, so easy to show, but difficult to describe and classify. An example of such an inspection can be the inspection of the surface of rubber hoses: verification of the correctness of prints and finding defects in the structure. Circular snakes are often covered in spiral interweaving patterns that will look different depending on the portion of the snake being examined and the current inspection site, and printing defects can come in a variety of forms: from spilled, fuzzy patterns to elements positioned at the wrong distances and locations.

Such defects are not very repeatable and each has its own unique shape, size or color, and all this, together with the high variability of the product itself, creates a large set of parameters and unknowns, which can be an insurmountable barrier for traditional algorithms, or even a classic case for system applications intelligent.

Intelligent vision systems and their real applications – a bull’s eye or an excess of form over content?

Recently, the subject of artificial intelligence has been a real hit – everything around us is promoted as “smart” or “intelligent”, and the visions and possibilities of the proposed solutions seem to be really promising. This phenomenon is also observed in the world of automation and robotics, and more specifically in the solutions offered for companies and factories.

At the moment, intelligent vision systems are not a cure for every problem and will prove themselves in solving a specific group of problems. Their implementation and implementation still require the involvement of a considerable amount of time of an experienced integrator, and in many applications classic algorithms will still be a better choice.

One of the challenges for the implementation of DL-based solutions is to create a set of data / photos, described with appropriate metadata, on the basis of which algorithms will learn how to make decisions. The dataset should be prepared carefully as it will be the main source of “knowledge” for the algorithms.

On the other hand, this approach allows us to solve problems that are out of reach for classical algorithms.

Undoubtedly, intelligent systems are a topic that arouses great interest among companies and factories, because their potential is huge.

The problem of frequent changes to the product – can it be solved with AI?

The automotive industry is the largest market for vision systems, where they are mainly used to control the correctness of assembly, i.e. checking the presence and position of components, reading codes and quality control of components. Defects of elements arise at the stage of production and assembly of components, and their specification and classification is a real challenge, because scratches, abrasions, shortcomings or dents can manifest themselves in any way, so their scope is often ambiguous and wide.

For automatic verification of such cases, one of the outputs could be a vision system based on deep learning methods that can cope with large scope and variability. In order to be able to analyze products with high variability, it is necessary to create sets of photos containing various arrangements and variants of the subject of inspection: sets of correct, incorrect photos, showing the subject of inspection in various positions, variants, arrangements, sets with incomplete objects, in different lighting, often rotated and distorted. Such collections can contain from hundreds to thousands of photos, and the more the algorithm learns, the more accurate it will be during inspection.

Summary.

With DL, some traditional machine vision techniques previously used have become less relevant, but knowledge is never out of date and there is always something to learn from each generation of innovation. It should be remembered that many of the problems of machine vision can be solved just as effectively using traditional programming methods, and often this is even more desirable than using DL. Artificial intelligence is certainly a powerful tool in the hands of programmers, but to fully use its potential, it takes a little more time to adapt to its considerable requirements.

Read more about our machine vision solutions (MVS).

Leave A Comment