Search papers, labs, and topics across Lattice.
This paper addresses the challenge of low-resource layout classification by introducing a novel CNN-based classifier that leverages layout-preserving augmentations to enhance generalization. By curating a complex-layout dataset and employing techniques like narrow anisotropic Gaussian masking and reflection-induced label transformations, the authors effectively mitigate the impact of data scarcity and semantic complexity in page layouts. The findings reveal that these layout-specific augmentations significantly boost classification performance, even in the presence of limited annotated data.
Layout-specific augmentations can dramatically enhance classification accuracy in low-resource settings, overcoming the challenges posed by complex page layouts.
Many digitized corpora suffer from low resources because annotations may be scarce, page scans are noisy and of poor resolution, or layouts are structurally complex in ways that negatively affect the quality of automatic transcription. Developing robust classification models for low-resource languages is inhibited by the lack of large-scale annotated data and by the frequent semantic complexity of page layouts. To this end, we have curated a complex-layout dataset, manually classified into eight distinct layout types based on their separator regions. To overcome data scarcity, we propose a novel training strategy in the form of a CNN-based classifier that employs strong, domain-aware augmentations to improve generalization. We utilize narrow anisotropic Gaussian masking to suppress incidental textual details while preserving essential separations, compelling the model to learn global geometric arrangements. Additionally, we implement reflection-induced label transformations to enrich the training distribution while maintaining label consistency across asymmetric categories. The results demonstrate that layout-specific augmentations can substantially improve page-level layout classification under severe annotation scarcity.
