Document Layout Analysis Using Detection Transformers

In the O&G (Oil & Gas) industry, unstructured data sources such as technical reports on hydrocarbon production, daily drilling, well construction, etc. contain valuable information. This information however is conveyed through various formats such as tables, forms, text, figures, etc. Detecting these different entities in documents is essential for building a structured representation of the information within and for automated processing of documents at scale. Our work presents a document layout analysis workflow to detect/localize different entities based on a deep learning-based framework. The workflow comprises of a deep learning-based object-detection framework based on transformers to identify the spatial location of entities in a document page. The key elements of the object-detection pipeline include a residual network backbone for feature extraction and an encoder-decoder transformer based on the latest detection transformers (DETR) to predict object-bounding boxes and category labels. The object detection is formulated as a direct set prediction task using bipartite matching while also eliminating conventional operations like anchor box generation and non-maximal suppression. The availability of sufficient publicly available document layout data sets that incorporate the artifacts observed in historical O&G technical reports is often a major challenge. We attempt to address this challenge by using a novel training data augmentation methodology. The dense occurrence of elements in a page can often introduce uncertainties resulting in bounding boxes cutting through text content. We adopt a bounding box post-processing methodology to refine the bounding box coordinates to minimize undercuts. The proposed document layout analysis pipeline was trained to detect entity types such as headings, text blocks, tables, forms, and images/charts in a document page. A wide range of pages from lithology, stratigraphy, drilling, and field development reports were used for model training. The reports also included a considerable number of historical scanned reports. The trained object-detection model was evaluated on a test data set prepared from the O&G reports. DETR demonstrated superior performance when compared with the Mask R-CNN on our dataset.

Combining Visual and Textual Features for Semantic Segmentation of Historical Newspapers

The massive amounts of digitized historical documents acquired over the last decades naturally lend themselves to automatic processing and exploration. Research work seeking to automatically process facsimiles and extract information thereby are multiplying with, as a first essential step, document layout analysis. If the identification and categorization of segments of interest in document images have seen significant progress over the last years thanks to deep learning techniques, many challenges remain with, among others, the use of finer-grained segmentation typologies and the consideration of complex, heterogeneous documents such as historical newspapers. Besides, most approaches consider visual features only, ignoring textual signal. In this context, we introduce a multimodal approach for the semantic segmentation of historical newspapers that combines visual and textual features. Based on a series of experiments on diachronic Swiss and Luxembourgish newspapers, we investigate, among others, the predictive power of visual and textual features and their capacity to generalize across time and sources. Results show consistent improvement of multimodal models in comparison to a strong visual baseline, as well as better robustness to high material variance.