Geometry-aware and autonomous reading-order learning for OCR-free document parsing

Purpose The purpose of this study is to address the task of visually rich document understanding, which extracts structured semantics from complex multimodal document images. Core challenges, including feature distortion from geometric deformations and structural deviations from complex layouts, jointly undermine extraction precision and robustness – the former due to inadequate cross-modal adaptation to local distortions and the latter due to constrained global topology modeling. Design/methodology/approach To address these limitations, the Geometry-Topology Collaborative Parsing Framework is proposed. This framework achieves robust document parsing through dual technical approaches. First, deformable kernels are used to correct geometrically warped text features, with channel attention mechanisms integrated for feature noise suppression. Second, cross-attention mechanisms and the XY-Cut algorithm work collaboratively to construct reading orders adapted to document topological structures. This co-design ensures simultaneous optimization of text feature enhancement and reading-order construction. Findings The framework demonstrates superior performance in document classification and structured information extraction through synergistic optimization of geometric correction and structural relationship decoding, significantly reducing reliance on error-prone optical character recognition (OCR) preprocessing. Originality/value This work establishes a novel paradigm for joint spatial-semantic document understanding via three key advancements: adaptive distortion rectification through deformable-convolutional text localization; density-aware reading sequence generation for robust layout parsing; and unified representation learning bridging visual features with structural topology without intermediate OCR dependencies.