A Clustering Backed Deep Learning Approach for Document Layout Analysis

Large organizations generate documents and records on a daily basis, often to such an extent that processing them manually becomes unduly time consuming. Because of this, automated processing systems for documents are desirable, as they would reduce the time spent handling them. Unfortunately, documents are often not designed to be machine-readable, so parsing them is a difficult problem. Image segmentation techniques and deep-learning architectures have been proposed as a solution to this, but have difficulty retaining accuracy when page layouts are especially dense. This leads to the possibilities of data being duplicated, lost, or inaccurate during retrieval. We propose a way of refining these segmentations, using a clustering based approach that can be easily combined with existing rules based refinements. We show that on a financial document corpus of 2675 pages, when using DBSCAN, this method is capable of significantly increasing the accuracy of existing deep-learning methods for image segmentation. This improves the reliability of the results in the context of automatic document analysis.