Machine Learning Homogeneity Inspector for detection of defects on surfaces

Machine or Deep Learning methods are an actual revolution on the market for image processing. It is a new method, which allows us to face challenges to the systems of image recognition, which are difficult to solve or cannot be solved at all with classic image processing. But there are applications which are better solved with classic methods.

Here is a short overview, to help decide which method is more effective for your application:

Machine Learming / Deep Learning

Classic image processing methods

Typical application

Typical application

surface inspection (cracks, scratches)

dimensional check

food-, plants-, wood-inspection

code reading

plastics, die casting mold

object detection

inspection of textile

robot guidance

imaging in medicine

print inspection

Typical characteristics

Typical characteristics

flexible objects

inflexible objects

variable orientation

fixed orientation

The customer has to give limited specifications, which contains correct and incorrect elements

The customer has to give a formalized specification with tolerance ranges

Advantages of Machine Learning

The advantages of Machine Learning are very clear. With Machine Learning you can backed by data:

  • learn unknown connections
  • modelizing processes
  • realizing adaptive mechanisms, which make the plant flexible and easily convertible

Modern production plants are nowadays highly complex. Processes are cross-linked with each other, machines, interfaces, and component parts are communicating with each other. Such industrial plants are predestinated for the optimization by methods of machine learning. Because they make it possible to make a prediction according to a large amount of data.

Challenges of Machine Learning

There are still several challenges to master. This would be first of all the high demands on the computing power of the CPU. Not every computing unit has enough power to be suitable for machine learning. Some Deep Learning applications demand souped-up hardware.

The second challenge is the required data set and the training effort. Concerning the training effort, it is what most Machine Learning applications require for gaining a classificator and therefore require a more or less large amount on manually classified training data. And here occurs a new problem, because in the industrial environment there are often a lot of good-part-examples, but only a few bad-(or error)-part-examples. Some errors even occur very rarely but are highly relevant.

The solution here is the so called one-class-classificatior. In this case you only train good-parts, of which are always plenty available.

For the training of the one-class-classificator the representation of the data is carried out with the help of the so called k-means-clustercenter. Because even in case of good-parts there are different feature characteristics, you should choose an over representation of data, this means the number of cluster centers is usually too high. Subsequent to that the cluster with a low number of representatives are deleted, to avoid the so called overfitting and also to consider, that there are not only representatives of good-parts, but also that there can be bad-parts contained in the random sample.

The Homogeneity Inspector

The EVT Homogeneity Inspector incorporates the complex functionality into an uncomplicated EyeVision software command, where the user has easy graphic tools available. Therefore the training process is really easy.

The Homogeneity Inspector parts the expected images in small individual segments and due to the previously defined error possibility it finds good-parts or errors in each of the segments. Subsequently it issues the size of the error in mm² and the sum of the total area of the error in the image.

Metal working and automotive: metallic surfaces

One application is the detection of surface defects and another to classify the error in the image in „scratch“ or „crack“.

Therefore even with metallic surfaces two characteristics have to be differentiated and treated algorithmic:

  1. typical structure and composition of the surface
  2. large-area characteristics, such as color or structural homogeneity

The homogeneity Inspector is able to detect and classify surface defects in the metal working, the automotive industry or in all areas of metal processing.

Because even slight aberrations in the car finish or other metallic surfaces lead to drawbacks of the visual overall impression. Often this is the reason for complaints or customer dissatisfaction.

Similar defect detection methods with metal is also suitable for weld seams or corrosion detection. A further application would be the detection of scratches on cars for example of rental cars.

Fiber board or other textured surfaces

With the surface inspection of mineral fiber boards or other structured surfaces, one has to find the error in the pattern. Also here a one-class-classificator is the best solution as there are not enough bad-parts. In case a new design or pattern has to be trained, one uses a large as possible sample (approx. 20 – 200 pieces) of the current production of this design as a representative average.

Also here two characteristics have to be differentiated and treated algorithmic:

  1. First characteristic: needling or texture in its typical shape and structure
    The needling of a fiber board is full of contrast and very easy to segment. The pinpricks should be geometrically similar if it is a good-part and geometrically noticable if it is a bad-part. Features such as: plain, roundness, axial ratio, average gray value, etc. are calculated.
  2. Second characteristic: large-area characteristics, such as e.g. the color application, the structural homogeneity, the existence or lack of structure, its distribution, etc. The whole board is segmented in overlapping windows of representative but fixed sizes. For each of these windows the software calculates histograms, which also contain the grey value distribution and the ratio of needling and background.

The surface and design defects that the Homogeneity Inspector can find in the production are for example: discoloration, blur of color and aberrations in the pattern such as holes, clefts, etc.

Spread the love

Leave a Reply

Your email address will not be published. Required fields are marked *