YOLO-DLA: A YOLO-based unified framework for multi-scale document layout analysis

Abstract

Document layout analysis (DLA) serves as a cornerstone of modern information systems, enabling efficient data extraction and structured knowledge organization. However, processing multi-scale layout documents has remained a bottleneck in the development of universal DLA frameworks. To address this limitation, we introduce the first multi-scale DLA dataset and propose a novel YOLO-based detection framework. Specifically: (1) To address the lack of fine-grained layout annotations in existing datasets, we construct AcadLayout, a specialized dataset for scientific documents with 13 layout element types (e.g., multi-level headings, formula, figure caption). (2) To address the challenges of multi-scale feature extraction, particularly for micro-scale elements, we innovatively incorporate the KWConv dynamic convolution method. (3) To achieve robust feature fusion across scales, we propose the PRDM-neck module, which uniquely integrates axial attention with multi-scale context aggregation. (4) To address scale imbalance in scientific documents, we propose a scale-aware curriculum learning strategy that progressively trains models from macro- to micro-scale elements (macro$\to$medium$\to$micro), effectively balancing detection performance across all scales.