Why Deep Learning
In the case of complex task problems, traditional programming using algorithmic descriptions of every possible variant is markedly more complicated than deep learning, making classification tasks significantly easier to solve than with existing algorithmic methods. Neural networks are particularly well qualified for many other challenges, such as for reflective surfaces, poorly illuminated environments, moving objects, robotics, and 3D, to name but a few, where conventional methods clearly reach their own limits and can be realized only at great cost and with great expertise.
Real-time classification of defects on metallic surfaces
Deep learning makes it possible to learn several defect classes at once. Vis-à-vis changing system environments such as light, contamination and aging, deep learning is algorithmically less conspicuous. Furthermore, classical algorithms are suited for cases involving exact determination of an object’s or defect’s position within an image, dimensional inspection, code reading, or post processing. Deep learning, in contrast, proves its value with very high reliability in identification rates. It will improve the quality of current image processing systems as well as open up new applications.