| Kontakt | Tel +49.711.23853.0 | info@AIT.de |

Deep Learning in Machine Vision

How does Deep Learning work in machine vision?

To do this, we first need to take a closer look at the known different control methods in the manufacturing process, quality inspection and traceability. A distinction is made between human visual inspections, conventional machine vision and machine vision with Deep Learning. But what exactly are the differences, the advantages and disadvantages of the individual methods?

Traditional Machine Vision vs Deep Learning

Deep Learning for complex situations

Deep Learning programmes function similarly to the human brain. They “imitate” the brain in their deep neural networks. Images, texts and numbers are used to train these neural networks. Differences in the data series are recognised and irregularities and changes are permanently searched for. Deep Learning delivers the decisive advantage over conventional machine vision in complex situations, with many exceptions and a high error library, high speed and reliability. This is precisely where conventional machine vision reaches its limits.

CLASSIFICATION

In order to distinguish between different object types in production, to identify defect types and even to check images, a selection of labelled images are taught in. This allows products to be identified and classified into classes. This is done based on their common characteristics such as colour, texture, material, packaging and defect type.

DETECT ANOMALIES

Deep Learning is used to quickly detect defects on complex parts and surfaces. This is done by learning a selection of good and bad images with marked defects, so that the normal deviation of the parts is learned and meanwhile a comprehensive understanding of the defects is developed.

LOCATE PARTS

In manufacturing, complex features and objects in a field of view become detectable with specialised deep learning image analysis software. The software detects features on choppy backgrounds, in poorly lit environments, on low-contrast parts and even on parts that bend or change shape. It locates parts despite variations in perspective, orientation, brightness, gloss and colour by learning from sample images.

CHALLENGING OCR

Even the most difficult OCR tasks in a manufacturing environment, such as needle embossed codes on metal parts, embossed characters on injection moulded products, laser etched codes on electronic components, label based codes on packaging, low contrast characters and codes on uneven backgrounds can be solved with artificial intelligence.

Classical machine vision and Deep Learning. The perfect complement.

But even where conventional machine vision fails and Deep Learning is used, both technologies can work together and complement each other:

DETECT POSITION AND DETERMINE POSITION

Components fall randomly onto a vibrating plate and are separated there. The camera system must detect the parts that are lying the right way round and output the position. Since the inspection objects lie irregularly and their features are very small, rule-based machine vision reaches its limits here. With Deep Learning we recognise the correct position of the parts and with classical machine vision the exact position is then determined so that a robot can grab the parts.

From idea to concept. The project evaluation.

Deep Learning Project Evaluation

As a manufacturer-independent component and solution provider, we have the most diverse products for the most diverse applications in our portfolio.