As part of the CSE Early Research Scholars Program, our group developed BASEDNet, a Convolutional Neural Network that scores baseline predictions for historical documents on a holistic basis. The network itself is implemented as a Binary Image Classifier, and it acts as a discriminator between 'good' and 'bad' baseline predictions. The model itself is defined and trained within src/train.py. Further, with the trained model, we also use a gradient-based decoder to optimize 'bad' baseline predictions with respect to two scores -- the holistic goodness score produced by BASEDNet, and another goodness score produced by dhSegment, a model that generates baseline predictions for historical documents.
Note: The data required for training our model is not included in the repo currently.
Details on how to use our model can be found here
This model was designed and developed by:
We received advising on this project from: