Complex Layout Classification in the Wild: A Low-Resource Approach with Layout-Preserving Augmentations

A new study proposes a low-resource method for classifying complex document layouts, using a convolutional neural network trained with domain-specific augmentations to overcome severe annotation scarcity [1]. The work, submitted to the arXiv preprint server on 15 June 2026, addresses a persistent obstacle in digitization projects: many corpora lack large-scale annotated data, and page scans are often noisy, low-resolution, or structurally intricate [1][2]. Sharva Gogawale, the paper’s sole listed author, curated a dataset manually sorted into eight distinct layout types defined by their separator regions [1][2]. To train a classifier on this limited material, the study introduces a CNN-based strategy that relies on two augmentation techniques. Narrow anisotropic Gaussian masking suppresses incidental text details while preserving essential separations, forcing the model to learn global geometric arrangements [1][2]. Reflection-induced label transformations then expand the training distribution while keeping labels consistent across asymmetric categories [1][2]. The results indicate that these layout-specific augmentations can substantially improve page-level classification when annotations are scarce [1][2]. The paper appears on arXiv, an open-access repository that hosts e-prints across mathematics, physics, computer science, and related fields [6]. Founded in 1991, arXiv passed the two-million-article milestone by the end of 2021 and now receives roughly 24,000 submissions per month [6]. The platform’s Labs framework, which supports community-built tools such as Bibliographic Explorer and CORE Recommender, allows third-party collaborators to develop experimental features that appear on article abstract pages [4][5]. arXiv states that collaborators have access only to minimal, anonymized user data and that any other use is prohibited without written consent [4]. The new layout-classification preprint is accompanied by links to code and data services, including Papers with Code and Hugging Face, which are integrated into the abstract page under the Code, Data and Media section [1][5]. The study does not report results on specific low-resource languages, but the abstract frames the work as a response to challenges in digitizing corpora where annotations are sparse and layouts are semantically complex [1][2]. The submission file size is listed as 36,012 KB [1].