MVTec officially released Halcon 17.12 Deep Learning Function

Already released Alcon 17.12 officially met with you in December 2017. It also provides two versions for you to learn!

One is called progress edition evolution, and the other is steady edition stable edition. Of course, anyone can guess who updates the fastest maintenance here! At present, the version that has been issued is the process version, which feels little changed compared with the previous version of Halcon 13, and has some more in-depth learning functions!

First, in-depth study:

With the release of 17.12, users will be able to train their own cellular neural classifiers (Convolutional Neural Networks for short, CNN), after which CNN can be used for new data classification.

1. training CNN

In Halcon, a CNN is trained by providing enough labeled training images. For example, in order to distinguish the samples showing scratches or contamination from good samples, all three types of training images must be provided: the images showing scratches must be labeled "scratches", and the images showing some contamination must be labeled "contaminated". The image showing a good sample must be in the "OK" category.

2. Use training networks

Once the network learns to distinguish a given class, for example, to determine whether the image shows scratches, contamination or good samples, the network can be put into operation. This means that users can apply the newly created CNN classifier to the new image data, and then match the classifier with the classes learned in training.

Typical areas of application for in-depth learning include defect classification (e.g., circuit boards, bottle nozzles or pills) or object classification (e.g., identification of plant species from an image).

2. Detection of specular reflection

Examining specular reflective surfaces poses a particular challenge because the observer cannot see the surface itself, but only the mirror of the environment. This poses serious problems for most surface detection methods, such as triangulation or shadow shape detection, because they usually rely on diffuse reflection.

HALCON 17.12 includes new operating instructions, which enable users to use deflection principle to detect defects in the mirror and part of the reflective surface. This method uses the mirror reflection mentioned above to observe the mirror image of the known pattern and the deformation of its surface.

3. Automatic Text Reader

HALCON 17.12 features an improved version of an automatic text reader that now detects and separates glued characters more strongly.

Surface Fusion of Multiple Three-Dimensional Point Clouds

HALCON now provides a new way to put multiple three-dimensional point clouds onto a sealed surface. This new method can combine the data of various 3D sensors, even from different types of data such as stereo camera, flight camera time and fringe projection. This technology is especially suitable for reverse engineering.

Improvement of hdevengine

The new HDevelop library export included in HALCON 17.12 is as easy and intuitive as calling any other C++ function, similar to calling HDevelop from a C++ program. The export of the new image processing library also supports Cmake projects, which can be easily configured for many popular IDE projects, such as Visual Studio.

Download address: https://www.51halcon.com/thread-954-1-1.html