PARL: Position-Aware Relation Learning Network for Document Layout Analysis

TL;DR

PARL introduces a vision-only approach to document layout analysis that eschews OCR-based fusion in favor of explicit structural modeling of the document's visual grammar. It simultaneously integrates BSP-DA for feature-level relational priors and GRC for decision-level relational reasoning, achieving state-of-the-art results on M6Doc, DocLayNet, and D4LA with significantly fewer parameters than multimodal rivals. The framework demonstrates strong efficiency (65M parameters) and robustness by leveraging spatial priors and graph-based context, reducing reliance on OCR accuracy. This work highlights the practical viability of structure-aware visual modeling for complex layouts, offering a path toward robust, real-time document understanding without text recognition bottlenecks.

Abstract

Document layout analysis aims to detect and categorize structural elements (e.g., titles, tables, figures) in scanned or digital documents. Popular methods often rely on high-quality Optical Character Recognition (OCR) to merge visual features with extracted text. This dependency introduces two major drawbacks: propagation of text recognition errors and substantial computational overhead, limiting the robustness and practical applicability of multimodal approaches. In contrast to the prevailing multimodal trend, we argue that effective layout analysis depends not on text-visual fusion, but on a deep understanding of documents' intrinsic visual structure. To this end, we propose PARL (Position-Aware Relation Learning Network), a novel OCR-free, vision-only framework that models layout through positional sensitivity and relational structure. Specifically, we first introduce a Bidirectional Spatial Position-Guided Deformable Attention module to embed explicit positional dependencies among layout elements directly into visual features. Second, we design a Graph Refinement Classifier (GRC) to refine predictions by modeling contextual relationships through a dynamically constructed layout graph. Extensive experiments show PARL achieves state-of-the-art results. It establishes a new benchmark for vision-only methods on DocLayNet and, notably, surpasses even strong multimodal models on M6Doc. Crucially, PARL (65M) is highly efficient, using roughly four times fewer parameters than large multimodal models (256M), demonstrating that sophisticated visual structure modeling can be both more efficient and robust than multimodal fusion.

Figures (6)

• Figure 1: As illustrated, this figure shows the "visual grammar" of document layout: elements do not exist in isolation but are closely correlated with their context and their own spatial position.
• Figure 2: Overview of our proposed method. The BSP-DA module injects spatial priors into the attention mechanism. It learns from global spatial relationships (via FFN Spatial) and semantic context (via Self-Attention) to compute an "Offset $\Delta \boldsymbol{o}_i$ Guidance," which corrects the "Original" sampling points into a more precise "Result." The GRC module refines the final classification by modeling contextual relationships. It constructs an F-S graph based on Semantic and Spatial cues, and then uses this graph context (via GAT Layers and a Dynamic Classifier) to optimize the final class predictions.
• Figure 3: Convergence comparison of PARL and baseline on the M6Doc validation set.
• Figure 4: Distribution of the $\lambda_i$ gating coefficients on the M6Doc validation set. The wide-ranging distribution indicates that the gating mechanism is dynamically learning to balance priors and has not collapsed to 0 or 1.
• Figure 5: Qualitative comparison of different models on documents with complex layouts. From left to right: DINO (M2Doc), DFINE, and PARL. Our model demonstrates superior performance in handling intricate structures.