Using Bipartite Matching and Detection Transformers for Document Layout Analysis Transfer Learning

Business documents represent useful information which could benefit from automatic interpretation. The task of document layout analysis seeks to identify and localize semantic structures in documents. Contemporary techniques approach this as a strictly visual task. However, recent progress in Natural Language Processing (NLP) has enabled the incorporation of language information. Multimodal techniques have been proposed for the task of document layout analysis. These models make use of region based object detection techniques which require defining surrogate tasks such as region proposals and non-max suppression. This thesis presents LayoutLMDet, a multimodal layout analysis model. LayoutLMDet approaches object detection as a direct set prediction task as described in "End-to-End Object Detection with Transformers". Using bipartite matching, LayoutLMDet removes the need for surrogate tasks, simplifying implementation. Leveraging a pretrained transformer encoder, LayoutLMDet is able to achieve a mean average precision of 49.5 on the DocLayNet test dataset. A qualitative comparison of LayoutLMDets performance on the DocBank dataset highlights the impact of data selection.