Solutions K Technologien K AI / Deep Learning

AI / Deep Learning

Deep learning works according to the principle of “learning”. Once the software has been trained, it can precisely assign new images to the corresponding categories. Deep learning is used in situations where conventional machine vision is not sufficient. When objects are complex and cannot be described using rules.

Classification

During classification, the entire image is assigned to a class. The classes only need to be made known to the network by means of corresponding sample images. In this method, an image is assigned a confidence value for the class, which indicates the match rate with which the presented image belongs to this class.

Object detection

With the help of object detection, various trained classes can be recognized in an image and localized in space. An object found is assigned to a class and marked with an enclosing rectangle. Recognizing and distinguishing between different objects is also used for reading fonts and optical character recognition (OCR).

Segmentation

Segmentation can be seen as a kind of refinement of object detection. Like a more accurate version of object recognition. This method assigns a class to each individual pixel of the image, including the background. This allows you to segment a found object precisely.

Anomaly detection

Anomaly detection finds deviations from a good version. The system is trained with error-free objects. The system then recognizes defects (e.g. cracks, holes, scratches) in the image as deviations. Advantage: The system learns to find errors without knowing them beforehand.

aku.deepLearningCloud

In classic machine vision and its typical tasks, there are often limit values and parameters that are used to determine quality. These values are assigned upper and lower limits and thus provide reliable values in conjunction with calibration and validation over the service life of the systems.

In the field of deep learning, however, there are tasks that very often require a dynamic adjustment of the parameters. Due to the production process, products change their appearance, which influences image processing but has no effect on the actual product quality. Particularly in the area of visual inspection, which was previously carried out by the human eye, the system must be enabled to learn in the same way as the human brain. To do this, the neural network must be supplied with image information and retrained. The customer provides the corresponding image data via the aku.deepLearningCloud, which we additionally train and test as part of a service. The resulting update can be imported into the system by the customer independently or remotely. The system is therefore in a continuous optimization and learning process and meets the highest demands in terms of quality and productivity.

Nowadays, data protection is a top priority when exchanging digital information. The sensitive image data from your production and from your products must reach us via a secure communication channel. Which route you choose is up to you – you can be sure that we will get your application to its destination.

CASE STUDIES

You might also be interested in: