Deep learning is something we encounter more and more often. Almost every industry magazine, online magazine and many newsletters write about it. You too have probably heard or read about it.
Often it is only clear that Deep Learning exists and what successes can be generated with it. Maybe you are wondering whether you can also use Deep Learning and what possibilities there are for you. Who wouldn’t want to further improve their productivity and the quality of their products?
We would like to give you an overview of how Deep Learning works. Afterwards you will also find out whether it is possible to use it in your company.
Deep Learning belongs to Machine Learning
In short, Deep Learning is a part of Machine Learning. Machine Learning is a generic term for machine pattern recognition. Similarities are determined from different data sets using an algorithm. However, for many dimensions of the data or if too little data is available, pattern recognition with this algorithm is often problematic.
Deep learning systems consist of so-called neural networks. They can cover several dimensions simultaneously and are ideal for the recognition and classification of objects. In these applications several features should be recognized and assigned in fractions of seconds.
These artificial neural networks are modelled on the human brain, which consists of natural neural networks.
Among other things, the many different characteristics are referred to as dimensions. For example, an apple has at least the shape, colour and stem as a dimension.
How does Deep Learning work?
Deep Learning consists of different possible networks for pattern recognition and classification. In the machine vision field, the concept of Deep Convolutional Networks (DCN) is usually used. The DCN is best suited for image analysis.
With such a DCN high-dimensional recognition tasks become possible. The network for classification is preceded by a network for dimensional reduction, because not all dimensions are always necessary for product recognition. This network reads the image data pixel by pixel in certain sectors that overlap each other. Only then is the data classified in another network.
This is what a DCN looks like in theory:

The Deep Learning network thus recognizes the product from the available image material. It can also learn the various features itself. However, a conventional machine learning system relies on data that you have entered manually.
However, the Deep Learning model must also be trained. This can take a lot of time. Without the right hardware, it can take days or even weeks. By using a GPU, the effort can be reduced to a few hours, because a GPU can calculate several processes at the same time.
How can you apply Deep Learning?
Deep Learning is especially helpful if you want to perform complex recognition and classification tasks. If you are already using machine learning, but sometimes reach the limits, Deep Learning may be the right approach for you.
Deep learning is particularly suitable for medical applications or complex products, as there are often several dimensions involved.
In the medical and life sciences sector, for example, Deep Learning can be used to automatically detect certain cells very easily. This is why Deep Learning is already used in cancer research.
Deep Learning is also used in quality control. Within the scope of automation, products that do not meet your quality standards can be sorted out.
What do you need for the implementation?
Successful implementation of deep learning for pattern recognition and classification requires appropriate hardware and deep learning software.
The hardware consists of a vision system and a graphics processor. This GPU is best located on a special frame grabber that you can install in almost any PC. The vision system can then be connected to the PC through the frame grabber, while most of the computing work is already done in the camera and frame grabber. This frame grabber is available separately or the CXP 12 frame grabber is only available in combination with the Basler boost.
If you would like to create an even more compact and cost-effective system, we recommend Embedded Deep Learning. This system is based on a combination of board-level cameras and embedded processing units. This ensures low unit costs and fast operational readiness.
We would be pleased if we could help you to exploit your potential even more. Please contact us if you have any further questions.