# Serey OCR: An Open Research Journal on Khmer Text Recognition and Document Intelligence

**Version:** 0.3 — October 21, 2026  
**Status:** Living technical research journal; draft for author review; **not peer reviewed**  
**Scope:** Khmer text recognition, document understanding, model efficiency, and reliability  
**Research identity:** Serey OCR / realserey

## Abstract

Khmer optical character recognition presents intertwined problems of visual recognition, Unicode representation, document layout, and offline deployment. This journal documents a research journey grounded in an existing Khmer OCR system, Serey OCR, and surveys advances that may inform future work: SVTRv2 and CTC recognition, self-supervised representation learning, Khmer-aware tokenization, vision-language document models, knowledge distillation, quantization, and uncertainty-aware hybrid OCR. It separates published findings from project observations and research hypotheses. The project has functioning offline recognition and experimental document-structure tooling, but the available real-document assessments are small, development-oriented, and not independently qualified; no broad Khmer accuracy claim is made. The goal is to identify scientifically interesting questions and credible opportunities for future evaluation, not to claim a new state-of-the-art model.

> **Publication boundary:** This public journal contains no private source PDFs, crops, personal information, government document contents, local paths, evaluation receipts, or unpublished model weights. Private reports remain local and are not offered as independently reproducible public benchmarks.

## 1. Motivation

Many general OCR benchmarks emphasize Latin or Chinese scene text, while practical Khmer recognition involves a script with consonant subscripts, dependent vowels, marks, complex shaping, and Unicode order that is not identical to visual spatial order. Beyond accurate line recognition, real documents demand reliable region discovery, reading order, field preservation, and structured extraction. These concerns create meaningful trade-offs between specialist recognition systems and large multimodal document models.

The Serey OCR project provides a concrete point of reference: a frozen specialist Khmer recognizer and a separate experimental document-analysis pathway. The journal asks what contemporary research can teach us about Khmer recognition rather than assuming that newer, larger models necessarily perform better.

## 2. Research journey to date

| Period | Theme | What was learned | Status |
|---|---|---|---|
| Early 2026 | Foundational Khmer OCR and local inference | Converting research models into usable local OCR involves decoding, preprocessing, export, and runtime constraints. | Engineering groundwork |
| September 2026 | Serey OCR baseline and multi-platform direction | A frozen baseline enables controlled research while keeping working mobile/desktop paths stable. | Baseline established |
| October 2026 | Recognition development experiments | Training configuration and data composition may yield improvements without changing the recognizer family. | Development only; no independent qualification |
| October 2026 | Source-grounded document examination | Detector omissions, truncated or misrecognized content, and reading-order errors are different failure classes. | Small source-inspected samples |
| October 2026 | Document layout technology survey | PaddleOCR, Docling, and Surya have distinct strengths, integration constraints, and licensing considerations. | Limited exploratory comparison; Surya weights not executed |
| October 2026 | Desktop structured-document experiments | Native document structure and manual review can preserve line identities and table relationships, but do not guarantee semantic correctness. | Experimental workflow |
| October 21, 2026 | Modern OCR architecture research | Khmer tokenization, SVTRv2 ablations, and self-supervised visual learning inform architecture hypotheses. | **Three literature studies; no Khmer model experiment** |

### What the development observations do—and do not—show

- **Detection:** At least one visually readable colored heading was absent from detected text-line regions. A recognizer cannot read a line it never receives.
- **Recognition:** A source-containing crop still produced an incomplete numeric sequence. The responsible stage has not been isolated.
- **Serialization:** In a small development study, order conventions contributed heavily to page-level edit distance. Diagnostic reordering is not an independently qualified accuracy improvement.
- **Structure:** A native document-layout framework recovered useful block and table geometry but still yielded unreliable ordering in some complex pages.
- **Workflow:** A complete PDF-to-edited-export workflow passed an engineering acceptance check for a local document. This does not establish transcription accuracy across document types.

The source materials for these observations are private. Ground-truth preparation in some experiments used machine-assisted draft transcription followed by owner corrections, creating possible anchoring bias. For this reason, the journal deliberately avoids publishing those diagnostic numbers as general OCR benchmarks.

## 3. Research questions

This journal is organized around six tracks rather than one proposed replacement model.

| Track | Architectural or scientific question | Published inspiration | Khmer-specific unknown |
|---|---|---|---|
| A — Specialist recognition | Can CTC-based SVTRv2 retain accuracy while reducing autoregressive inference cost? | SVTRv2: MSR, FRM, SGM [1] | Complex marks, long lines, and cluster-level alignment |
| B — Self-supervised vision | Can unlabeled text images improve the visual representation? | Linguistics-aware masked image modeling [2] | Preservation of subtle Khmer visual distinctions |
| C — Script-aware tokens | Should Khmer output labels be code points, grapheme clusters, orthographic syllables, or a lossless hybrid? | Unicode Khmer encoding and segmentation [3, 4] | Coverage, visual alignment, reversibility |
| D — Document VLMs | When can page-level models recover regions or structure missed by modular pipelines? | PaddleOCR-VL-1.6 [5], GLM-OCR [6], DeepSeek-OCR 2 [7] | Khmer fidelity, unsupported insertions, cost |
| E — Efficient inference | How do distillation and low precision affect sensitive Khmer distinctions? | Quantization tooling and recognition literature [8] | Error concentration in vowels, marks, and digits |
| F — Hybrid reliability | Can uncertainty identify which regions deserve independent verification? | Selective prediction and calibration literature | Calibration without hallucinated corrections |

## 4. Study 01 — Khmer tokenization and CTC feasibility

**[Read the standalone study](studies/01-khmer-tokenization-svtrv2.md).**

SVTRv2 offers a valuable contrast to the current specialist baseline. Its multi-size resizing (MSR) limits destructive aspect-ratio distortion; feature rearrangement (FRM) produces a CTC-aligned feature sequence from two-dimensional visual features; semantic guidance (SGM) supplies linguistic context during training but can be omitted at inference [1].

For a CTC target with `L` tokens and `R` adjacent repeated-token pairs, the input-time dimension must satisfy `T >= L + R` to permit an alignment. A Khmer cluster containing several code points may therefore require more CTC labels at codepoint granularity than at cluster granularity. That mathematical observation is not a benchmark result. Cluster tokenization introduces its own difficulty: vocabulary growth, rare clusters, and the need for reversible fallback.

**Provisional conclusion:** The suitability of SVTRv2 for Khmer cannot be determined without understanding the target segmentation and the recognizer's time-step budget. The first scientific questions concern Unicode fidelity, syllable segmentation, cluster coverage, and visual-sequence alignment, not which checkpoint is best.

## 4A. Study 02 — Anatomy of SVTRv2: MSR, FRM and SGM

**[Read the standalone study](studies/02-anatomy-of-svtrv2.md).**

Du et al. (ICCV 2025) show that SVTRv2's improvements result from several complementary changes: aspect-ratio-aware resizing, higher-resolution feature maps, local/grouped convolution with global attention, a feature rearrangement module (FRM) that constructs a CTC-suitable sequence from two-dimensional features, and semantic guidance (SGM) used only during training [1].

In the paper's **published English scene-text ablations**, MSR improves accuracy by 15.3 percentage points versus fixed 32×128 resizing in a narrow text aspect-ratio subgroup, FRM improves multi-oriented scene text by 2.46 points, and SGM improves the average occluded scene-text score by 5.11 points. These are *different subsets and distinct comparisons*, not Khmer results or additive effects. The original paper's output character vocabulary and mostly short scene-text tasks differ substantially from full Khmer administrative lines.

**Provisional conclusion:** More vertical feature detail is conceptually relevant to Khmer marks, while FRM and SGM raise important alignment and exact-transcription questions. None of the published ablations establishes that replacing Serey OCR's decoder with CTC will improve Khmer accuracy.

## 4B. Study 03 — Self-supervised visual representation learning for Khmer OCR

**[Read the standalone study](studies/03-self-supervised-visual-learning.md).**

The third study examines Masked Autoencoders (MAE), discriminative/generative pretraining (DiG), masked encoder–decoder pretraining (MaskOCR), and Linguistics-aware Masked Image Modeling (LMIM) [9–12]. The core distinction is that **unlabeled visual pretraining learns representations, not verified character transcriptions**. Downstream Khmer recognition and evidence-grounded evaluation still require suitable labels.

In LMIM's English Union14M seven-subset ablation, adding a visually transformed guidance branch raises reported mean accuracy from **77.7% to 80.6%**, while guidance plus global alignment raises it to **81.2%** [12, Table 1]. The result is not a Khmer improvement. The paper uses 80% random patch masking and a teacher-feature reconstruction target; its official released pretraining recipe uses an existing teacher checkpoint and illustrates eight-GPU distributed execution.

**Provisional conclusion:** More unlabeled text images are useful only insofar as the pretraining objective learns invariance to irrelevant appearance **without learning invariance to Khmer vowels, marks, and subscript distinctions**. The best masking geometry, augmentation invariances, pretrained visual features, and target-domain benefit remain unknown.

## 5. Related research areas

### 5.1 Self-supervised pretraining

Study 03 surveys MAE, DiG, MaskOCR and LMIM and distinguishes unlabeled visual pretraining from supervised or synthetic decoder training [9–12]. Khmer-safe visual invariance is the central unresolved issue; English/Chinese results do not establish Khmer transfer.

### 5.2 Vision-language document OCR

Modern document VLMs pursue broader tasks than line transcription. PaddleOCR-VL-1.6 emphasizes refinement of underperforming document regions [5]. GLM-OCR uses a compact multimodal architecture and multi-token prediction [6]. DeepSeek-OCR 2 proposes visual causal flow for semantic ordering of visual tokens [7]. Their contributions are noteworthy, but aggregate parsing scores are not substitutes for script-specific exact transcription evidence.

### 5.3 Efficiency and reliability

Distillation, quantization, and selective second-pass OCR could be useful only if they preserve or improve the particular distinctions that matter in Khmer. Average accuracy may hide small but high-impact failures involving dates, amounts, legal identifiers, and diacritics. Future evaluation should consider character- and cluster-level errors, field preservation, calibration, and latency together.

## 6. Evidence policy

Each journal claim is labeled conceptually as one of:

1. **Published result:** Supported by a named paper on its stated dataset and configuration.
2. **Project observation:** An engineering or source-inspected finding in a limited, local sample.
3. **Research hypothesis:** Plausible architectural transfer to Khmer that has not been validated.
4. **Open question:** The current evidence cannot distinguish competing explanations.

Private research material is never reclassified as independent public ground truth. The journal does not assert that SVTRv2 or any VLM currently surpasses the Serey OCR baseline on Khmer.

## 7. Research agenda

This is a *reading and analysis agenda*, not an implementation schedule.

| Order | Topic | Intended scholarly output |
|---|---|---|
| 01 | Khmer tokenization and CTC feasibility | Representation assumptions and alignment constraints — initial note complete |
| 02 | SVTRv2 internal modules | Architecture, ablations and Khmer transfer caveats — **literature study complete** |
| 03 | Self-supervised Khmer visual features | Critical survey of masking, contrastive learning, and augmentation risks — **literature study complete** |
| 04 | Document VLM architectures | Comparative examination of encoders, decoders, token compression, grounding |
| 05 | Distillation and quantization | Accuracy–latency–memory trade-off framework for offline devices |
| 06 | Hybrid OCR and uncertainty | Grounding, calibration, abstention, and selective verification |
| Synthesis | Architectural options for future Khmer OCR work | Hypothesis-led technical review; no predetermined implementation |

## 8. Limitations

This is a living research record, not a peer-reviewed paper. The limited Serey OCR observations are not a controlled comparison of published models on a public Khmer dataset. Model-size, runtime, and benchmark claims from external papers are not directly comparable across their hardware, settings, versions or evaluation corpora. The absence of a published Khmer score in the references does not prove a model is incapable of Khmer recognition. Stronger conclusions require independently grounded evidence.

## 9. References

1. Du, Y. et al. **SVTRv2: CTC Beats Encoder–Decoder Models in Scene Text Recognition.** ICCV 2025. https://arxiv.org/abs/2411.15858 ; code: https://github.com/Topdu/OpenOCR
2. Zhang, Y. et al. **Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition.** CVPR 2025. https://openaccess.thecvf.com/content/CVPR2025/html/Zhang_Linguistics-aware_Masked_Image_Modeling_for_Self-supervised_Scene_Text_Recognition_CVPR_2025_paper.html
3. Hosken, M., with contributors N. Lindenberg and M. Sok. **Khmer Encoding Structure.** Unicode Technical Note 61, revision dated October 2024. https://www.unicode.org/notes/tn61/utn61-Khmer_Encoding_Structure_V1.pdf
4. Unicode Consortium. **Unicode Standard Annex #29: Unicode Text Segmentation.** https://www.unicode.org/reports/tr29/
5. Zhang, Z. et al. **PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training.** 2026. https://arxiv.org/abs/2606.03264
6. Duan, S. et al. **GLM-OCR Technical Report.** 2026. https://arxiv.org/abs/2603.10910
7. Wei, H. et al. **DeepSeek-OCR 2: Visual Causal Flow.** 2026. https://arxiv.org/abs/2601.20552
8. ONNX Runtime. **Model quantization documentation.** https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html

9. He, K. et al. **Masked Autoencoders Are Scalable Vision Learners.** CVPR 2022. https://openaccess.thecvf.com/content/CVPR2022/html/He_Masked_Autoencoders_Are_Scalable_Vision_Learners_CVPR_2022_paper.html
10. Yang, M. et al. **Reading and Writing: Discriminative and Generative Modeling for Self-Supervised Text Recognition.** ACM Multimedia 2022. https://arxiv.org/abs/2207.00193
11. Lyu, P. et al. **MaskOCR: Scene Text Recognition with Masked Vision–Language Pre-training.** TMLR 2024. https://openreview.net/forum?id=KNAWoKKpi3
12. Zhang, Y. et al. **Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition.** CVPR 2025. https://arxiv.org/html/2503.18746 ; code: https://github.com/zhangyifei01/LMIM

---

**Citation suggestion:** *Serey OCR Research Journal*, v0.3 (October 21, 2026), living technical research notes. Published by realserey: https://huggingface.co/spaces/realserey/serey-ocr. Please do not cite this draft as a peer-reviewed finding.