Deep learning for smart manufacturing: Methods and applications

[suchona.kani.z]

The problem is that an expert has to determine each time which features are relevant, and these have to be designed in such a way that they allow a distinction to be made from other classes. This is a huge logistical and financial effort. With deep learning, this is not necessary, because the network takes care of feature extraction automatically and can train them "end to end". Quality control is carried out very simply using the label of the image, so it must be known in advance which image belongs to which class.


CNNs are a class of neural networks. They use convolutions to process images. These are networks that are resistant to shifts in an image. Filters detect properties that can then be used later for classification. Objects are therefore found in all positions.

The following image shows a simple example of a CNN. You can see that a small part of the image is selected again and again and that there are three layers. This is because it is an RGB image (i.e. with the three color channels red, green and blue). With the help of convolutions and pooling (information compression), the image is compressed greece consumer email list more and more (x/y direction) while the number of different convolutions (z direction) increases. This is referred to as a gain in context information while at the same time losing localization information. This happens in the area known as feature learning. A vector is then output with a probability for each class. The highest probability is then the class that the neural network "thinks" is the right one for the image.

Example for CNN
Source:

Another advantage of deep learning is so-called transfer learning. Neural networks are trained on a large data set of images. They are adapted to a specific use case using just a few images. An example of this is medical data, which usually has just a few 100 to 1000 examples. In practice, a network is pre-trained with a million images and later adapted using the few medical images. This method works so well because it primarily trains the network's feature recognition, making it gradually more complex.

Traditional computer vision methods are used when microcontrollers are necessary or when little sample data is available.

Applications of Computer Vision with Deep Learning
Autonomous driving
Autonomous driving would simply not be possible without computer vision. One example of an algorithm is Tesla Vision, the computer vision module of the Tesla autopilot. This is a deep learning approach that segments images and forwards information to an AI, which then makes decisions about driving behavior. It is important that the AI ​​recognizes road signs, other road users and road markings. All of this is no (insurmountable) problem thanks to computer vision with deep learning.