Train the Tool

Once a labeled instance of each character has been completed, it's time to train the tool. Press the Brain icon, and the software will begin its computation for the training.

The labeled instances will be used during the deep learning phase of the tool's processing via the neural network. It's not necessary to label multiple instances of each character, but the more labeled instances there are, the better the results will be during the training phase. Training will increase the accuracy of the tool. The more you train, the more you tune your neural network towards your particular training set. The tool starts with a very generalized model, and every time you train, the tool is going to key in more and more to what you have in your training set. So, if your training set isn't a very good representation of what you'll encounter at run time, you would be better off to minimize the amount of training you'll do.

If you have labeled everything, and the labeled features represent the expected appearances (and preferably involve the instances of each character), training will improve the performance of the tool. However, if you have an unrepresentative set of images (for example, characters on a different background or a slant that isn't expected at runtime), training may narrow the performance to the unrepresentative set of characters, so training isn't advised.

To have a successful implementation of the tool, you need to have a set of images that contain the full range of expected variations of the characters at runtime.

Important tool parameters involved in training:

  • Train Selection: Defines the percentage of labeled images that will be used during training. By default, 10% of the labeled images will be used during training; the other 90% are unlabeled images.
  • Keep checkbox: Defines that you will reuse the same labeled images during each subsequent training.
Tip: If you have images that you always want to be included in the training, after the View is accepted, you can right-click the image, and from the menu, expand the Train selection and select Always.
Note: If the tool is able to read the full image database of all of the characters in your images with 100% accuracy, you do not need to train.