Document Image Classification with Intra-Domain Transfer Learning and Stacked Generalization of Deep Convolutional Neural Networks

The area-based deep convolutional neural network framework for learning document architecture

is introduced. The contribution of this work includes effective training of classifiers on the basis

of region and the effective assembly of classification of document images.

The basic level is used to learn to "inter-domain" transfer by exporting weights from a pre-

trained VGG16 structure to an ImageNet dataset to train the document classifier on images of

entire documents

This is done by exploiting the nature of area-based effect modeling, a secondary level of "in-

field" transfer learning is used for the rapid training of deep learning models of image parts.

Finally, generalization stacked-based clustering is used to combine predictions of basic deep

neural network models.