Evaluation of Generic Deep Learning Building Blocks for Segmentation of 19th Century Documents

Abstract

Although the field of computer vision has grown tremendously due to the rise in popularity of convolutional neural networks, historical document analysis has seen a lackluster increase in research and development. Generic computer vision has reached the point where it can be used to outperform the existing, specialized tools for document analysis, as demonstrated by dhSegment using ResNet. We build upon this insight that generic models can produce state-of-the-art results by implementing, training, and evaluating other generic computer vision models on historical document segmentation tasks. We show that this unspecialized approach to document analysis is not limited to ResNet and that innovation in this domain can spawn from various general building blocks for computer vision.