Deep learning

Deep learning in image processing

In industrial production processes, conventional vision systems reach their limits in certain tasks that are not a problem for humans.
Despite the limitations and fluctuations of human sensory perception, human judgment is still the best solution here.

The solution: Intelligent image processing systems with self-learning algorithms – deep learning.

Deep learning: the next step in image processing

Our Deep Learning technology area uses leading technologies based on state-of-the-art machine learning algorithms (machine learning or deep learning).

This makes it possible to overcome challenges such as …

Data volume too large to scale
Tasks too complex to program
Non-existent or limited laws and rules
Irregular or unknown deviations from the norm
Reliable identification of natural products where there are no identical parts, only unique ones
Our DL technologies enable three basic “human” tasks to be solved fully automatically and highly efficiently:

Discover all kinds of qualitative anomalies in your components. For example, scratches on difficult or textured backgrounds, faulty assembly or weaving defects in textiles.
Our HMV technologies identify these and many other anomalies on the basis of an expected image that is learned using the permissible appearance of an object, including tolerated tolerances.

Detect and identify individual or multiple features in an image.
For example, highly deformed characters on a noisy background or complex objects in bulk.
Our HMV technologies locate and identify complex features and objects by learning from labeled images on which the target features are marked.

Classify objects or an entire scene.
For example, identify objects based on their appearance or packaging.
Or classify acceptable and unacceptable defects.
Our HMV technologies learn to reliably distinguish between the different classes based on labeled images.

How you benefit from deep learning

A key feature of deep learning is a self-learning system (deep learning) based on representative sample images.
This significantly improves recognition performance compared to conventional industrial image processing methods.
In addition, the development time and time-to-market of DL solutions are considerably shortened – with significant cost benefits.
Feasibility studies can also be carried out in hours instead of days.

Basic characteristics of deep learning:

Images of good samples are used to train a so-called expected image (permissible appearance, including tolerated and permissible tolerances).
The imaging source is irrelevant.
Images of the test specimens are evaluated against the expected image in the series test.
This means

  • No complex feature or defect libraries necessary, as new defects are not necessarily trained.
  • Teach-in of new products without great effort.
  • Increasing performance of the system with increasing representativeness of the sample of images (higher recognition performance and robustness against permissible scattering).
  • Significantly better results compared to the best human testers.
  • No software development necessary.
  • Solving challenging tasks that are otherwise impossible to program.

Deep learning in practice

Here you will find a selection of possible applications for our DL technologies:

Tasks and challenges in the inspection of machined workpieces

  • Typically, there are many different types with complex forms and manifestations
  • Different surfaces and surface properties, depending on tool quality and the quality of further processing, must be tolerated
  • Some errors are only visible under very specific combinations of lighting, camera and test specimen orientation

Screws

Inspection of quality defects (scratches, nicks, pressure marks, stains)

To inspect this type of screw, it is rotated around the horizontal axis.
The images are captured and processed directly in the HMV system.
The model used for processing and evaluation is based on a representative selection of images without defects, but with the permissible manufacturing scatter.

The model takes into account acceptable variations in surface texture and the very complex shape of the self-tapping screw.

After training, the system is ready for the test.
Errors on the test specimens can be identified and marked very quickly and reliably.

Carbide tools

Inspection of wear, damage or breakage (chipping)

To inspect carbide tools, they are rotated around the horizontal axis.
The images are recorded and processed and evaluated directly in the HMV system.
The model used for processing and evaluation is based on images of OK samples.
The model takes into account acceptable variations in surface texture and the very complex shape of the tool.

After training, the system is ready for the test.
Defects on the test specimens can be identified and marked very quickly and reliably.
As the model is trained on the basis of “good patterns”, unknown (new) defects are also recognized.

Tasks and challenges in the inspection of pad prints

  • Several prints applied in succession can lead to relative but acceptable shifts (registration problem)
  • Varying amounts of ink result in characters or lines that appear thicker or thinner
  • Random textures of the substrate (e.g. brushed or otherwise textured metal surfaces)

Dials

Inspection of the quality of the print (faulty, incomplete)

To inspect dials, images are taken and processed directly in the HMV system.
The model used for processing and evaluation is based on a representative selection of images of OK samples.

The model takes into account acceptable variations in print or surface texture as they occur in the selection of OK samples.

After training, the system is ready for testing.
Errors are quickly and reliably identified and marked.

Button

Inspection of the quality of the print (faulty, incomplete)

Images are captured and processed directly in the system to inspect buttons.
The underlying model for processing and evaluation is based on a representative selection of images of OK samples.

The model takes acceptable variations in pressure or surface texture into account.
After training, the system is ready for testing.
Defects are quickly and reliably identified and marked.

Tasks and challenges in the automatic inspection of electroluminescence images

  • Very large variation in luminescence from cell to cell or from module to module is technically expected and tolerated
  • Defects such as micro-cracks can be very fine and are difficult to see on the highly irregular background
  • A large number of different errors can occur.
    This makes it impossible to develop a simple and robust algorithm, reliably detect the errors and at the same time keep the proportion of pseudo errors low

Solar cells

Automatic inspection of solar cells using electroluminescence images (e.g. micro-cracks)

To inspect solar cells, electroluminescence images are captured and processed directly in the system.
The underlying model for processing is based on a representative selection of images without defects and on labeled images with defects.

Due to the highly structured or textured background, micro-cracks are the most difficult to detect.
The model learns to distinguish micro-cracks and other defects from similar appearing structures in the background.

After training, the system is ready for testing.
Errors on the cells can be identified and marked very quickly and reliably.

Tasks and challenges in the inspection of fabrics / woven fabrics / textiles

  • Highly complex fabric patterns and a large number of pattern variants.
    This rules out simple methods for inspection.
  • Extremely strongly varying optical appearance.
    Deformations due to the flexible structure of the fabric or other variations, e.g. the yarn diameter, influence the appearance of the fabric.
  • Defects in textiles can come in countless shapes and types.
    Searching for all defects is not an economical option.

Textiles

To inspect textiles, images are captured and processed directly in the system.
The underlying model for processing is based on a representative selection of images of good samples.
These represent weaving patterns, yarn properties, color characteristics and tolerated imperfections.
The model learns to distinguish faults from textiles with similar structures.

After training, the system is ready for testing.
Defects on the textiles can be identified and marked very quickly and reliably.

Tasks and challenges in the inspection of textured surfaces

  • There are many types and variants of surface texturing on a wide variety of materials
  • Production processes are designed in such a way that random variations are created, making each part unique
  • Errors can occur in countless forms, which not only manifest themselves through a local change in contrast, but also a local change in texture

Textured surfaces

Inspection of textured surfaces (scratches, knocks, pressure marks, intolerable damage)

To inspect textured surfaces, images are captured and processed directly in the system.
The model used for processing and evaluation is based on a representative selection of images without defects and on labeled images with defects.
The model learns to distinguish defects from similar appearing structures of the surfaces.

After training, the system is ready for testing.
Defects on the textured surfaces can be identified and marked very quickly and reliably.

Training

Basic training "Development ViDi models"

The aim is for participants to be able to independently adapt existing models to changes.

Contents:

Learn the basic options of the ViDi software (ViDi red, ViDi blue, ViDi green).
Customer-specific / application-specific topics, depending on the implementation

Duration: 2 days
Training location: i-mation Rottweil or at the customer’s premises
Number of participants: max. 3

The company

Get in touch with us

Would you like to find out more about the possible applications and benefits of deep learning?
Just give us a call or send us an e-mail.

info@i-mation.de
Tel.
+49 741 942 286-00

Cookie Consent with Real Cookie Banner