DeepLearning at HALCON

WE ARE STRUGGLING WITH THE DEEP LEARNING CAPABILITIES OF HALCON'S LATEST VERSION 18.11.

One of our users asked us, "I tried labeling with the HALCON Deep Learning Tool, but I don't know what to do after that.

The latest HALCON deep learning functions include image classification, object recognition, segmentation, and anomaly detection.

The HALCON Deep Learning Tool is a free tool for annotating (pre-training) for image classification and object recognition; it is not possible to train or actually apply the tool. To do this, you need to run the script in HALCON's development environment Hdevelop, you need a HALCON development license, you need to learn HALCON's scripting syntax, and you need to understand HALCON's deep learning style. It is almost impossible to say, "I just want to try it out."

WE HAVE THEREFORE DEVELOPED AN APPLICATION OF HALCON'S SAMPLE SCRIPTS FOR DEEP LEARNING, WHICH CAN BE USED FOR TRAINING AND PERFORMANCE EVALUATION WITH SIMPLE OPERATIONS. (AS OF MARCH 27, 2020, "IMAGE CLASSIFICATION" AND "ANOMALY DETECTION" ARE SUPPORTED.)

We are trying it out on some sample images right away, and it seems to be able to make a good or bad judgment of about 70 points after learning about 100 images. The system can detect major defects at a level that can be noticed at a glance, and tends to make mistakes in images that do not seem to differ much from the good ones. I am not sure if I will be able to get closer to 100 points if I continue learning or if I will not be able to make much progress. If you can try it, experience it, and decide that it is still OK, it could be a great weapon.