Artificial Neuronal Networks in Industrial Use
Today, artificial neuronal networks are written about a lot and are researched actively. Although the underlying theory dates back to the 1940s, the big boom only began a few years ago thanks to increasingly powerful processors. Since, neuronal networks have found their way into industry.
Picture 1: Automatically Detected License Plate on a Vehicle at Stettbacher Signal Processing AG.
Within so-called artificial intelligence, machine learning takes on a key function and solves problems for which purely algorithmic methods would be too complex. In contrary to classically programed algorithms, in machine learning the instructions are not converted into source code by the programmer, but a neuronal network is used or designed, for example, which is capable of learning a problem solution. This is explained in a simple example: Firstly, it should be automatically determined whether a given image shows a vehicle (passenger car, truck). If so, secondly, the location of the vehicle’s license plate should be determined. Finally, the text on the license plate should be identified. See Picture 1.
To solve the first subtask, Stettbacher Signal Processing AG uses a pre-trained network. Such networks, which have already learned how to solve various tasks and can recognize people, pets or vehicles in photos, for example, are freely available. The quality and reliability of these networks must be verified using the company’s own data (pictures). So now, if a vehicle is detected in the photo, the license plate should be located. For this second subtask, SSP has trained its own network. To do this, a large amount of photos had to be taken or collected first as training data. The photos contain license plates from different perspectives and under various lighting conditions. Since the geometry, color and appearance of vehicle license plates are relatively uniform, only a few hundred photos were needed. In other cases, thousands or more records are needed. Accordingly, the training would be very computationally expensive. Normally, this is run on specially equipped computers. The network trained by SSP localizes and recognizes license plates reliably. Finally, the third and last step is to read the license plate: For this, another pre-trained network is used, which can identify texts in pictures. The picture section with the previously rectified license plate is given to this network. And now the task is solved.
It’s always amazing what neuronal networks can deliver. But they are constructed very clearly. The so-called neurons form layers, which reweight and combine the inputs (e.g. picture pixels) or the exits of the previous layer. However, real networks are consistent of many millions of neurons. From a certain number of layers, it’s called deep learning.
SSP uses the technology successfully in customer projects in the fields of machine vision and robotics. There, some of the reoccurring tasks are detection and localization of complex objects, the determination of their location and orientation in the room, as well as the grasping and/or placing of the objects.