Page Stream Segmentation (PSS)

• Page Stream Segmentation (PSS) is the task of converting a continuous stream of scanned pages into distinct documents by detecting boundary cues.
• It leverages OCR, layout analysis, and visual features to address challenges like layout variability and class imbalance.
• Applications span archival digitization, enterprise processing, and comic segmentation, with methods evolving from CNNs to Transformer models.

Searching arXiv for the specified PSS papers to ground the article in current research metadata. Page Stream Segmentation (PSS) is the task of transforming a continuous stream of scanned pages into correctly delimited documents or other coherent page segments. In large-scale digitization and ingestion workflows, pages often arrive without explicit boundaries, so a PSS system must infer where a new unit begins while preserving document context. In the archival and enterprise setting, PSS is typically framed as per-page binary boundary detection over an ordered stream; in comics, it has also been formalized as sequence-aware multiclass labeling in which boundaries emerge from semantic page labels. Across these settings, the core technical issues are context modeling, OCR robustness, layout variability, class imbalance, and the choice between local page-pair decisions and full-stream sequence modeling (Wiedemann et al., 2017, Heidenreich et al., 2024, Ortega et al., 14 Jul 2025).

1. Problem formulation and scope

In document-processing pipelines, PSS addresses a concrete operational problem: documents are batch-scanned as a continuous stream of single page images without explicit document boundaries, yet downstream systems require contiguous multi-page units. A standard formulation uses a boundary indicator sequence $y_1,\dots,y_N$ with $y_i \in \{0,1\}$, where $y_i = 1$ denotes the beginning of a new document and $y_i = 0$ denotes continuation of the same document. One formal statement defines the input as a sequence of $N$ pages $P = (p_1, p_2, \dots, p_N)$ and the output as a partition into $M$ contiguous documents $D = (d_1,\dots,d_M)$, where each $d_k = (p_i,p_{i+1},\dots,p_j)$ is a contiguous subsequence (Heidenreich et al., 2024).

This binary boundary-detection view is explicit in both archival and enterprise-oriented work. The 2017 multimodal CNN study formulates PSS over scanned page images $I_1,\dots,I_N$ and OCR-derived texts $y_i \in \{0,1\}$0, with $y_i \in \{0,1\}$1 labeled as the beginning of a new document (ND) and $y_i \in \{0,1\}$2 as same document (SD). Although that system is not an explicit sequence model such as an HMM or CRF, it incorporates predecessor information to capture local stream context (Wiedemann et al., 2017). The 2024 LLM study adopts an even more localized streaming formulation, specifically a pairwise setup with $y_i \in \{0,1\}$3 and $y_i \in \{0,1\}$4, so each decision considers $y_i \in \{0,1\}$5. This is aligned with a streaming constraint: predictions must be made sequentially without global hindsight (Heidenreich et al., 2024).

A broader formulation appears in comic-book analysis. There, PSS is defined as partitioning a scanned comic book into coherent semantic sections by labeling each page in an ordered stream $y_i \in \{0,1\}$6. The segmentation function $y_i \in \{0,1\}$7 assigns each page a class from $y_i \in \{0,1\}$8, and segment boundaries are induced by transitions in the label sequence or by the explicit boundary label First-Page (Ortega et al., 14 Jul 2025). This suggests that PSS is not intrinsically limited to binary document starts; it can also serve as a sequence-structuring problem in which semantic page types and boundary recovery are jointly modeled.

2. Benchmarks, corpora, and data regimes

Research on PSS has been strongly shaped by dataset availability. One recurring constraint is that realistic, public benchmarks have been scarce, with much prior work relying on private datasets or small corpora that do not reflect production conditions. The benchmark landscape described across recent work spans archival collections, tobacco-industry documents, and comic books.

Corpus	Domain and size	Distinguishing characteristics
Archive22k	22,741 pages from 100 binders	German federal archive; heterogeneous layouts, fonts, occasional handwriting, variable OCR quality
Tobacco800	1,290 document images	Truth Tobacco Industry Documents; letters, invoices, hand-written documents
TABME++	~110,000 streams; ~1.2 million documents; ~3.3 million pages	English tobacco archive benchmark with Microsoft OCR and preserved TABME splits
CoSMo corpus	430 classic comic books; over 20,800 pages	Public-domain comics with five page classes and pronounced class imbalance

Archive22k was developed in a German federal archive project in which roughly one million pages spanning 1922–2010 were batch-scanned; the experimental subset contains 100 binders, with 80 binders and 17,376 pages for training and 20 binders and 5,095 pages for testing. Its heterogeneity is central: layouts, fonts, page ages, handwriting, and OCR quality all vary substantially (Wiedemann et al., 2017). Tobacco800, by contrast, is a smaller public evaluation set of 1,290 document images from the Truth Tobacco Industry Documents; it includes letters, invoices, and hand-written documents, and is described as smaller and more homogeneous in corporate design than Archive22k, while still heterogeneous in layout and content over time (Wiedemann et al., 2017).

The 2024 paper introduces TABME++, an enhanced version of the English-language TABME benchmark drawn from the Truth Tobacco Industry Documents archive. TABME++ preserves TABME’s train, validation, and test splits and its synthetic stream construction, but replaces Tesseract OCR with Microsoft OCR. Its aggregate statistics are much larger than those of earlier public benchmarks: approximately 110,000 streams, approximately 1.2 million total documents with approximately 44,800 unique documents, approximately 3.3 million total pages with approximately 122,500 unique pages, and approximately 9.5 billion GPT-2-tokenized tokens with approximately 349.7 million unique tokens. The average pages per document are approximately $y_i \in \{0,1\}$9. Compared with the original TABME, blank pages are reduced from 2,785 (2.27%) to 466 (0.38%) (Heidenreich et al., 2024).

The comic domain introduces a different label ecology. The CoSMo corpus consists of 430 classic comic books from the public-domain Digital Comic Museum, totaling over 20,800 pages. Pages are labeled as Cover, Advertisement, Text Story, Story, or First-Page. The class distribution is markedly imbalanced: Story pages account for 71%, while Cover is 2.4%, Advertisement 8.8%, First-Page 13.4%, and Text Story 4.2% (Ortega et al., 14 Jul 2025). This differs qualitatively from binary enterprise PSS because boundaries are intertwined with semantic macro-structure.

3. Modeling paradigms

PSS methods in the cited literature progress from feature-engineered text baselines and multimodal CNN fusion to parameter-efficient decoder LLM fine-tuning and full-stream Transformer encoders.

The 2017 study uses a layered multimodal design. Its text baseline is a linear SVM over sparse OCR-derived text features, with ablations over unigrams, LDA topic composition, topic difference between consecutive pages, and copied predecessor-page features. In parallel, it introduces a text-CNN with a 300-dimensional randomly initialized embedding layer, a 1D convolution with 350 filters of kernel size 3, global max pooling, a dense layer with 256 ReLU units, dropout $y_i = 1$0, and a sigmoid output. The image branch uses a VGG-16 backbone pretrained on ImageNet, with original pretrained layers frozen, a new dense layer with 256 units and dropout $y_i = 1$1, and a sigmoid output. Fusion is performed by training the text-CNN and image-CNN separately, removing their final sigmoid layers, augmenting text features with LDA topic proportions and topic-distance features, concatenating a three-page text window, and then passing the combined vector $y_i = 1$2 through an MLP with a dense 256-unit ReLU layer, L2 regularization factor 0.01, dropout $y_i = 1$3, and a sigmoid classifier. The decision equations are reported as

$y_i = 1$4

with boundary prediction at threshold $y_i = 1$5 (Wiedemann et al., 2017).

The 2024 study evaluates two model families on TABME++. The encoder baselines are RoBERTa (text-only), DiT (vision-only), LiLT (text + layout), and LayoutLMv3 (text + vision + layout); all are fine-tuned end-to-end on page pairs. Because encoder context length is limited to $y_i = 1$6 tokens, text models encode the first $y_i = 1$7 and last $y_i = 1$8 tokens and concatenate them. The decoder family consists of Mistral-7B and Phi-3-mini, adapted with parameter-efficient fine-tuning via LoRA on a single H100 GPU with 4-bit quantization through Unsloth and BF16 LoRA weights. The training objective is next-token prediction over the completion, not the instruction, with structured JSON classification outputs (Heidenreich et al., 2024).

For decoder inference, the input is serialized into two explicit sections, Prior text and Page text, each containing OCR-extracted 2D text for the corresponding page. The prompt requests a JSON object of the form {"label": 1} for a new-document start or {"label": 0} otherwise; if the model returns malformed JSON or the wrong key, the system defaults to {"label": 0}. This makes the output interface deterministic and stream-compatible (Heidenreich et al., 2024).

The 2025 CoSMo model shifts from page-pair classification to whole-stream sequence modeling. It is an encoder-only Transformer with two main variants. The vision-only variant uses frozen SigLIP visual embeddings and a single token per page. The multimodal variant adds OCR extracted with Qwen2.5-VL-32B and text embeddings from Qwen3Embedding-0.6B. Each modality is projected to a shared 768-dimensional space via a three-layer MLP with GELU, LayerNorm, and dropout 0.4; the projected vectors are L2-normalized. The multimodal sequence is constructed by multi-token interleaving, $y_i = 1$9, with absolute positional encodings. A 4-layer Transformer encoder with 4 heads per layer, hidden size 256, input embedding size 768, feedforward dimension 3072, and dropout 0.4 contextualizes the entire stream. A three-layer MLP classifier then outputs a 5-way distribution per page (Ortega et al., 14 Jul 2025).

Across these systems, explicit sequence post-processing is notably absent. The 2017 multimodal CNN uses no CRF or HMM; the 2024 LLM pipeline reports no smoothing heuristics, minimum or maximum document-length constraints, or beam search; and CoSMo likewise applies no CRF, HMM, smoothing, or post-processing (Wiedemann et al., 2017, Heidenreich et al., 2024, Ortega et al., 14 Jul 2025).

4. OCR quality, modality choice, and contextual information

OCR quality is a central variable in PSS because many boundary cues are textual: titles, salutations, closings, identifiers, and page-type markers. The strongest direct ablation on OCR quality appears in TABME++, where the only change between TABME and TABME++ is the OCR source. Microsoft OCR improves text availability and fidelity relative to Tesseract, uncovers previously missed titles and identifiers, and reduces blank pages from 2.27% to 0.38%. On this basis, Mistral-7B fine-tuned on TABME++ reaches page-level F1 $y_i = 0$0 and document-level F1 $y_i = 0$1, whereas on original TABME its page-level F1 is $y_i = 0$2 and its MNDD rises to $y_i = 0$3 from $y_i = 0$4. RoBERTa recall is also markedly higher on TABME++ than on TABME (Heidenreich et al., 2024).

The representation of OCR output also matters. TABME++ uses Microsoft OCR via Azure Computer Vision OCR and serializes the extracted text to preserve 2D layout structure on the page. The paper states that this 2D projection is beneficial for models that exploit layout cues. No bespoke reading-order normalization beyond the OCR-provided structure is discussed (Heidenreich et al., 2024). In the comic setting, structured OCR from Qwen2.5-VL-32B provides reading order and text types such as Titles, Panels, and Content Text, which are described as critical for Text Stories and Advertisements (Ortega et al., 14 Jul 2025).

The empirical relation between modality and performance is more nuanced than a simple “more modalities are better” rule. In the 2017 archival study, combining text and image features improves over single-modality systems on both Archive22k and Tobacco800, and the model explicitly treats visual and textual cues as complementary under OCR noise (Wiedemann et al., 2017). In the 2024 benchmark, however, DiT, a vision-only encoder, is the strongest among the encoders, while LayoutLMv3 and LiLT underperform it, and LiLT underperforms RoBERTa. The paper therefore states that added modalities are not a guaranteed win in PSS unless the architecture and training regimes are well aligned (Heidenreich et al., 2024). In comics, visual features dominate macro-structure: vision-only CoSMo is already very strong, while multimodality yields consistent but modest gains and is most useful in resolving ambiguous cases such as First-Page versus Advertisement or prose-like Text Stories (Ortega et al., 14 Jul 2025).

Context modeling follows a similar pattern. Earlier enterprise PSS work uses predecessor windows: the 2017 method copies predecessor features and concatenates a three-page text window, while the 2024 LLM system uses previous page plus current page. CoSMo extends context to the entire book via self-attention over the whole page stream. This suggests a progression from local continuity cues to global contextualization, especially in domains where boundaries depend on broader narrative structure rather than only adjacent-page transitions (Wiedemann et al., 2017, Heidenreich et al., 2024, Ortega et al., 14 Jul 2025).

5. Evaluation metrics and reported performance

PSS evaluation varies with formulation, but the core distinction is between page-level correctness and operationally meaningful segmentation quality. For binary boundary detection, the 2024 work defines $y_i = 0$5 and $y_i = 0$6 as the predicted and ground-truth boundary page indices, with

$y_i = 0$7

It then reports

$y_i = 0$8

At the document level, exact segment matches are evaluated with analogous precision, recall, and F1. Stream-level metrics are MNDD, the minimum number of drag-and-drops required to correct the predicted segmentation, and STP, the fraction of streams perfectly segmented (Heidenreich et al., 2024).

The 2017 study instead emphasizes per-page Accuracy and Cohen’s $y_i = 0$9, where

$N$0

Its use of $N$1 reflects the uneven SD/ND class distribution (Wiedemann et al., 2017). In comics, CoSMo adopts single-page Macro-F1, document-level F1 over segments under an IoU threshold, Panoptic Quality with $N$2, and a stream-level MnDD measure (Ortega et al., 14 Jul 2025).

Study and setting	Best reported model	Reported performance
Archive22k	Multimodal MLP (Image + Text)	0.929 accuracy, $N$3
Tobacco800	Multimodal MLP (Image + Text)	0.911 accuracy, $N$4
TABME++	Mistral-7B-FT	Page F1 0.987, Document F1 0.967, MNDD 0.811, STP 0.800
Comic books	Multimodal CoSMo	98.10 F1-Macro, 95.08 PQ, 0.437 MnDD

The result patterns are instructive. In the 2017 archival study, multimodal fusion outperforms both text-only and image-only models on Archive22k and Tobacco800, with strong gains in $N$5, which the paper treats as especially relevant under class imbalance (Wiedemann et al., 2017). On TABME++, fine-tuned decoder LLMs substantially outperform smaller multimodal encoders and a bag-of-words XGBoost baseline. Mistral-7B-FT reaches page precision/recall/F1 of $N$6 and document precision/recall/F1 of $N$7, while Phi-3-mini-FT reaches page F1 $N$8 and document F1 $N$9. Zero-shot LLMs are unreliable: Mistral-7B-ZS has page F1 $P = (p_1, p_2, \dots, p_N)$0, Phi-3-mini-ZS page F1 $P = (p_1, p_2, \dots, p_N)$1, and GPT-4o-ZS on a 10% test subset page F1 $P = (p_1, p_2, \dots, p_N)$2, all far behind fine-tuned decoders (Heidenreich et al., 2024).

In comics, CoSMo reports that larger general-purpose vision-LLMs are weaker than the task-specific encoder. Qwen2.5-VL-32B zero-shot reaches 87.12 F1 with a CLS prompt and 88.26 with an OCR-style prompt, whereas single-page multimodal CoSMo reaches 97.82 F1-Macro and 98.41 accuracy. On full-book sequence modeling, vision-only CoSMo with SigLIP reaches 97.30 F1-Macro, 94.50 PQ, and 0.632 MnDD, while multimodal CoSMo with multi-token interleaving improves to 98.10 F1-Macro, 95.08 PQ, and 0.437 MnDD (Ortega et al., 14 Jul 2025).

6. Operational pipelines, limitations, and future directions

Operational PSS systems are shaped by streaming requirements, OCR cost, hardware constraints, and correction effort. In the enterprise-oriented LLM pipeline, preprocessing uses commercial OCR with 2D text layout preserved in serialization, and segmentation proceeds sequentially: for each page, OCR the previous and current page, serialize them into a prompt, obtain a JSON label from the model, default to continuation on invalid output, and start a new document only when the label is 1. The paper reports that a single H100 suffices for fine-tuning 4–7B decoders with LoRA and 4-bit quantization, and emphasizes sample efficiency, with strong validation metrics after approximately 5,000 updates and early stopping around 20,000 updates, corresponding to approximately 0.297 observed epochs (Heidenreich et al., 2024).

The 2017 multimodal CNN pipeline is similarly deployment-oriented but uses lighter local context. Pages are OCRed, images are OTSU-binarized and resized to $P = (p_1, p_2, \dots, p_N)$3, text is tokenized and stemmed with digits replaced by #, LDA topic proportions and topic distances are computed, text-CNN and image-CNN representations are extracted, and an MLP decides whether each page begins a new document. The paper notes that this scales to large page streams because pages are processed independently with lightweight predecessor context and no global sequence inference (Wiedemann et al., 2017).

The comic-book setting has different practical trade-offs. CoSMo processes entire books, extracting a frozen SigLIP embedding for each page and, optionally, OCR plus text embeddings. This is trained and inferred on an NVIDIA L40s GPU. The paper characterizes the model as lightweight relative to large VLMs, but does not report runtime, memory footprint, or throughput (Ortega et al., 14 Jul 2025).

Several limitations recur across the literature. Domain coverage remains narrow: TABME++ is English and tobacco-industry-centric, Archive22k is German archival material, and CoSMo focuses on Golden Age Western comics. Privacy constrains realistic public document benchmarks because operational PSS data often contain PII or sensitive corporate information. Synthetic stream construction with replacement can introduce duplication artifacts and class-imbalance shifts, making deduplication strategies consequential. OCR dependency remains substantial even when vision helps; in comics, stylized fonts and complex layouts limit textual reliability, and in enterprise PSS, OCR quality strongly affects downstream recall and document metrics (Heidenreich et al., 2024, Wiedemann et al., 2017, Ortega et al., 14 Jul 2025).

Future directions stated in the cited work include domain-adaptive OCR and fine-tuning for broader domains such as healthcare, finance, and multilingual corpora; more effective integration of vision and layout into decoder pipelines; wider windowing beyond page pairs; stronger modeling of long-range dependencies; layout-aware and cross-lingual extensions; and handling of special cases such as double-page spreads, which CoSMo does not model as a distinct unit (Heidenreich et al., 2024, Ortega et al., 14 Jul 2025). A plausible implication is that PSS is evolving from local binary boundary classification toward domain-specific sequence modeling in which OCR, layout, and visual macro-structure are treated as complementary but not uniformly beneficial signals.