Christoph Wagner, Product Manager Embedded Vision at MVTec Software, comments.

What is the difference between "machine learning" and "deep learning"?

Christoph Wagner: In industrial image processing, machine learning and deep learning can be used to replicate and even surpass the performance of the human sense of sight and the understanding of visual information. With the help of deep learning, as a sub-area of machine learning, complex correlations can be learned and trained using examples.

Another technology within machine learning is "super vector machines", for example. In contrast to deep learning, features have to be defined and verified manually. In deep learning, a neural network takes over this manual step: features are identified and extracted independently and automatically as part of the training process.

What possibilities does deep learning open up for image processing technology and for inspection based on image processing?

Christoph Wagner: Deep learning can be used for typical classification applications such as defect detection or differentiating between "good" and "bad" objects. With an appropriately trained network, all image processing tasks can be solved in which a system decides whether certain types of defects are present in the image or not. Deep learning is particularly suitable for applications where conventional machine vision methods do not help. This is the case, for example, when the complexity of objects or defect classes means that no generally valid approach can be found to solve the problem.

Due to the variety of potential defects, it is almost impossible to manually develop algorithms that can detect and localize the entire spectrum of conceivable defects. These can be a wide variety of damage such as scratches, cracks or dents, which in turn can have many different shapes or sizes. Using conventional methods for defect detection, machine vision experts would have to view and evaluate a large number of images individually and use them to program an algorithm that describes the defect in as much detail as possible. This process would be very time-consuming and expensive.

Deep learning can significantly simplify defect detection and inspection: The technology independently learns specific defect characteristics on the basis of which certain problem classes can be identified. If the user uses pre-trained deep learning networks, such as those included in MVTec's Halcon image processing software, only a few hundred sample images are required for each class to be taught. The algorithms can use the sample images to train and then reliably recognize a wide variety of defect types.

What are the limits of deep learning in image processing technology and image processing-based inspection?

Christoph Wagner: Deep learning is typically used in the fields of classification, object detection and semantic segmentation.

As a complementary technology, deep learning is a useful tool for supporting machine vision applications. However, such complex tasks cannot usually be solved with deep learning alone, as the technology is merely another method for classifying data. A comprehensive toolbox that offers all image processing tools is of great importance in this context. This is because complex applications, including pre- and post-processing, can only be fully mastered through a combination of different methods.

Other factors need to be taken into account, especially in embedded vision systems: Because large amounts of data have to be analyzed here, training usually devours a lot of resources and is therefore unprofitable on embedded devices. The classification of data also requires a great deal of effort, so that large computing capacities and corresponding hardware are necessary. A standard CPU is therefore usually not sufficient, especially for extremely time-critical applications. Instead, powerful hardware such as a high-performance GPU is required, which tends to be the exception in industrial embedded systems.

What are the algorithms required for deep learning as part of the image processing software capable of?

Christoph Wagner: As part of a comprehensive training process, the deep learning algorithms can independently learn certain patterns that are typical for corresponding features. The system analyzes pre-categorized images, automatically assigns them to a specific class and checks whether this "prediction" corresponds to the actual category. This process is repeated until an optimal "prediction" result is achieved. In this way, models (classifiers) can be trained with which newly captured images can be categorized into the classes learned here.

Which applications for image processing technology with deep learning already exist today, and which are conceivable for the future?

Christoph Wagner: There are a large number of deep learning architectures, each of which has certain advantages and disadvantages for the respective applications. For robust detection rates in highly specialized and complex applications, customers need a network that is optimized and trained for their requirements. The principle of "one size fits it all" is usually not relevant here.

Many applications based on deep learning are used in the embedded sector. A wide variety of hardware components are used here to accelerate the runtimes of deep learning algorithms. This requires a framework that can be used on a wide range of dedicated deep learning computing units.

Will MVTec be presenting any new deep learning products at the embedded world trade fair?

Christoph Wagner: Halcon 18.11 offers outstanding new deep learning features, such as semantic segmentation and object detection, which also achieve impressive runtimes on embedded devices. Visitors to embedded world can experience this technology up close during a live demonstration at the MVTec stand. An AGX Xavier board from Nvidia with Halcon will analyze various objects in real time.

MVTec at embedded world 2019: Hall 4, Stand 203