NetLay: Layout Classification Dataset for Enhancing Layout Analysis

Abstract

Within the domain of historical document image analysis, the process of identifying the spatial structure of a document image is an essential step in many document processing tasks, such as optical character recognition and information extraction. Advancements in layout analysis promise to enhance efficiency and accuracy using specialized models tailored to distinct layouts. We introduce NetLay, a new dataset for benchmarking layout classification algorithms for historical works. It consists of over 1,300 images of pages of printed Hebrew (or Hebrew‑character) books in a variety of styles, categorized into four different classes based on their layout (the number of text columns and regions). Ground truth was crafted manually at the page level. Furthermore, we conduct an in‑depth performance evaluation of various layout classification algorithms, which are based on deep‑learning models that learn to extract spatial features from images. We evaluate our algorithms on NetLay and achieve state‑of‑the‑art results on the task of layout classification for historical books.