Familiarization with Text Recognition using Recurrent Neural Networks. The accuracy of the networks is not state of the art, but the main objective was to create a minimal size network for reasonable results. The number of parameters for the final version has about 85K parameters.
Images from the generated dataset. For a generation, only the vanilla MNIST dataset was used.
Ground truth strings were encoded following the rules established in this work:
- Numbers now have the prefix "s_"
- Empty space is encoded with the symbol "-"
- Space between numbers is encoded as "|"
Prediction of the network on the test dataset is saved in an HTML file using the Jinja2 template. Here is an example of the saved file from the Results directory:
