Two Level Document Image Classification

Abstract

Classifying documents is an important process for organizations that are responsible for keeping a large number of documents in a digital archive. In this paper, a two-level method was used to classify approximately 253 class of documents. In the first stage, the documents were visually classified. In the second stage, unclassifiable documents were tried to be classified using natural language processing (NLP). Training set was created by classifying approximately 28.000 documents by hand. Studies were carried out on single-page documents. Performances were measured by making experiments on the full, 1/2 and 1/3 of document using AlexNet for image classification, bag-of-words and LSTM algorithm for NLP. The performances of the study were measured according to different options. The success rate was 75% and 85% at the end of first and second stages respectively