The MLP classifier in HALCON has always been known for excellent classification results. In HALCON 12 it has been enhanced in several ways:
Training can now be done using a single class.
Confidence values provide a real indication of “how well” a sample can be classified.
Training is significantly faster.
It is possible to identify irrelevant features during the training process.
In defect inspection applications, it is often impossible to predict what kind of defects will occur. In this case you can now train the classifier using only good samples and HALCON will automatically generate a rejection class. The new operator set_rejection_params_class_mlp has been introduced for this purpose.
In addition, the MLP classifier now returns confidence values that provide an indication as to whether a sample is similar to two or several classes. This would be an indication that it cannot be identified reliably and you can label that sample as “unknown”, if the confidence values are too similar. The process that makes this possible is called regularisation and it can be enabled with the new operator, set_regularization_params_class_mlp. Regularisation smoothes the boundaries between classes, and thus the confidence values provide a better indication as to whether or not a sample can be classified reliably.
The automatic text reader tool in HALCON 12 uses the classifier in the segmentation process. The automatic generation of rejection samples as well as regularisation make this possible.
Last but not least, the MLP training is now parallelised internally and can be several times faster on processors with multiple cores.
Need to learn more ? Contact support@multipix