Deep Neural Networks for Page Stream Segmentation and Classification

In this manuscript we propose a novel method for jointly page stream segmentation and multi-page document classification. The end goal is to classify a stream of pages as belonging to different classes of documents. We take advantage of the recent state-of-the-art results achieved using deep architectures in related fields such as document image classification, and we adopt similar models to obtain satisfying classification accuracies and a low computational complexity. Our contribution is twofold: first, the extraction of visual features from the processed documents is automatically performed by the chosen Convolutional Neural Network; second, the predictions of the same network are further refined using an additional deep model which processes them in a classic sliding-window manner to help finding and solving classification errors committed by the first network. The proposed pipeline has been evaluated on a publicly available dataset composed of more than half a million multi-page documents collected by an on-line loan comparison company, showing excellent results and high efficiency.