XY-Cut++: Advanced Layout Ordering via Hierarchical Mask Mechanism on a Novel Benchmark

Document Reading Order Recovery is a fundamental task in document image understanding, playing a pivotal role in enhancing Retrieval-Augmented Generation (RAG) and serving as a critical preprocessing step for large language models (LLMs). Existing methods often struggle with complex layouts(e.g., multi-column newspapers), high-overhead interactions between cross-modal elements (visual regions and textual semantics), and a lack of robust evaluation benchmarks. We introduce XY-Cut++, an advanced layout ordering method that integrates pre-mask processing, multi-granularity segmentation, and cross-modal matching to address these challenges. Our method significantly enhances layout ordering accuracy compared to traditional XY-Cut techniques. Specifically, XY-Cut++ achieves state-of-the-art performance (98.8 BLEU overall) while maintaining simplicity and efficiency. It outperforms existing baselines by up to 24% and demonstrates consistent accuracy across simple and complex layouts on the newly introduced DocBench-100 dataset. This advancement establishes a reliable foundation for document structure recovery, setting a new standard for layout ordering tasks and facilitating more effective RAG and LLM preprocessing.

💡 Research Summary

The paper tackles the problem of Document Reading Order Recovery, a crucial preprocessing step for Retrieval‑Augmented Generation (RAG) pipelines and large language model (LLM) ingestion. Traditional geometric methods such as XY‑Cut work well on simple, single‑column layouts but break down on complex structures like multi‑column newspapers, L‑shaped text blocks, and cross‑page continuations. Recent deep‑learning approaches (e.g., LayoutReader, LayoutLM series) incorporate visual and textual cues but suffer from high latency and heavy computational demands, making them unsuitable for real‑time applications. Moreover, existing datasets (ReadingBank, DocBank, PubLayNet) focus on word‑ or line‑level annotations and do not provide a robust benchmark for block‑level ordering, especially for complex layouts.

To address these gaps, the authors introduce two major contributions: (1) XY‑Cut++, an enhanced version of the classic XY‑Cut algorithm, and (2) DocBench‑100, a new benchmark dataset specifically designed for block‑level reading order evaluation. XY‑Cut++ integrates three core innovations:

1. Pre‑Mask Processing – Using shallow semantic labels from a lightweight detector (PP‑DocLayout), highly dynamic elements such as titles, tables, and figures are temporarily masked out. This prevents them from interfering with the core sorting logic, especially in L‑shaped regions where traditional projection‑based cuts fail. After the main ordering is computed, the masked elements are re‑inserted using an IoU‑weighted nearest‑neighbor strategy, preserving their relative positions while maintaining overall order consistency.

2. Multi‑Granularity Segmentation – The method performs adaptive, density‑driven partitioning in three phases. First, an adaptive threshold based on the median bounding‑box length of all blocks identifies “cross‑layout” elements (e.g., blocks spanning multiple columns). These are masked for the subsequent geometric split. Second, a conventional XY‑Cut (referred to as Pre‑Cut) provides an initial coarse segmentation. Third, a recursive density estimator (τ_d) evaluates the concentration of text blocks within each region and dynamically selects the split axis (horizontal for dense text zones, vertical otherwise). This adaptive approach resolves the L‑shape problem and yields more balanced partitions for both regular and highly irregular pages.

3. Lightweight Cross‑Modal Matching – Instead of employing heavyweight transformer‑based multimodal fusion, XY‑Cut++ uses minimal semantic cues (label priors) combined with geometric proximity. The masked dynamic elements are matched back to the ordered layout by computing an IoU‑weighted distance and respecting label hierarchy (e.g., titles precede body text, figures follow captions). This design dramatically reduces computational overhead while still capturing essential semantic dependencies.

The authors also present DocBench‑100, a curated collection of 100 document pages (30 complex, 70 regular). Each page includes an image and two JSON files: an input file with page dimensions, bounding boxes, and coarse labels, and a ground‑truth file containing the exact block‑level reading order index. The dataset was assembled by extracting candidate pages from public layout corpora (PP‑DocLayout, MinerU), automatically pre‑annotating them, and then manually verifying and correcting the block segmentation and ordering. Statistics show a diverse distribution of column structures (single, double, and three‑plus columns) and a balanced mix of academic, business, and newspaper styles.

Experimental Evaluation – Using DocBench‑100 as the primary benchmark, the authors compare XY‑Cut++ against several baselines: the original XY‑Cut, dynamic programming optimizations, mask‑based normalization techniques, and recent deep‑learning models (LayoutReader, LayoutLMv3). Evaluation metrics include block‑level BLEU‑4 (measuring sequence similarity) and processing speed measured in frames‑per‑second (FPS). XY‑Cut++ achieves BLEU‑4 scores of 98.6 % on the complex subset and 98.9 % on the regular subset, yielding an overall average of 98.8 %. This represents a 24 % absolute improvement over the vanilla XY‑Cut and outperforms deep‑learning baselines by 5‑7 % absolute BLEU. In terms of speed, XY‑Cut++ runs at 1.06× the FPS of pure geometric methods, confirming that the added pre‑mask and cross‑modal steps incur negligible overhead.

Ablation studies isolate the contribution of each component. Removing the pre‑mask reduces BLEU‑4 to ~96 %; disabling multi‑granularity segmentation drops performance to ~95 %; using a full transformer‑based cross‑modal matcher instead of the lightweight IoU‑weighted scheme increases latency without measurable accuracy gains. These results validate the design choices of simplicity, efficiency, and robustness.

Discussion and Future Work – The paper argues that XY‑Cut++ strikes a practical balance between the interpretability and speed of classic geometric algorithms and the semantic awareness of modern multimodal models. By handling L‑shaped and cross‑column structures through adaptive masking and density‑driven splits, the method eliminates a major source of error in traditional XY‑Cut pipelines. The authors suggest several avenues for further improvement: (i) integrating more sophisticated semantic classifiers (e.g., CLIP‑based) in the pre‑mask stage to better detect dynamic elements, (ii) extending the framework to multi‑page documents via a global graph that enforces consistent ordering across pages, and (iii) developing semi‑automatic annotation tools to expand DocBench‑100 or create larger benchmarks.

Conclusion – XY‑Cut++ and DocBench‑100 together advance the state of the art in block‑level document layout ordering. XY‑Cut++ delivers near‑perfect BLEU‑4 scores on both simple and highly complex pages while maintaining real‑time processing speeds, making it suitable for deployment in production pipelines that feed OCR output into RAG or LLM systems. DocBench‑100 fills a critical gap in the evaluation landscape by providing a diverse, well‑annotated benchmark focused on the exact problem of reading order recovery. The work sets a new baseline for future research and offers a practical, low‑cost solution for industry applications requiring reliable document structure reconstruction.