PDF EXTRACTION PIPELINE - RESEARCH REPORT

Ngày: 20/03/2026

---

TỔNG HỢP 10 ĐỐI THỦ

---

1. AMAZON TEXTRACT

STEP	CÔNG NGHỆ
OCR Engine	Amazon Textract OCR (proprietary ML-based)
Layout Analysis	Deep learning - tự động detect paragraphs, columns, forms. Nhận diện reading order tự động
Table Extraction	Table detection với row/column boundaries. Hỗ trợ merged cells nhất định
Data Validation	Confidence scores cho từng field. Queries API cho phép extract specific fields
Output Format	JSON (Sync/Async) - Blocks, Lines, Words structure. Table output riêng biệt

---

2. GOOGLE DOCUMENT AI

STEP	CÔNG NGHỆ
OCR Engine	Google Vision OCR (20+ năm research)
Layout Analysis	Enterprise OCR - detect blocks, paragraphs, lines, words, symbols. Deep learning based
Table Extraction	Form Parser - extract tables, KVPs, selection marks. Hỗ trợ simple tables (ko merged cells)
Data Validation	Confidence scores, bounding boxes cho từng element
Output Format	JSON với entities, pages, blocks. Custom Extractor (Generative AI)

---

3. AZURE FORM RECOGNIZER (Document Intelligence)

STEP	CÔNG NGHỆ
OCR Engine	Azure AI Vision (proprietary)
Layout Analysis	Deep learning models - extract text, tables, KVPs, selection marks
Table Extraction	Layout API - extract table structure (rows, columns). Hỗ trợ complex tables
Data Validation	Confidence scores, accuracy reports. Build custom models
Output Format	JSON (analyzeDocument response). Prebuilt models cho invoices, receipts, etc.

---

4. ADOBE PDF EXTRACT API

STEP	CÔNG NGHỆ
OCR Engine	Adobe Sensei AI/ML (proprietary)
Layout Analysis	Deep learning - detect headings, lists, paragraphs, footnotes, columns, reading order
Table Extraction	Extract tables với cell data, headers, properties. Output CSV/XLSX option
Data Validation	Confidence scores, bounding box coordinates
Output Format	JSON (comprehensive) + optional CSV/XLSX cho tables + PNG cho images

---

5. ABBYY FLEXICAPTURE

STEP	CÔNG NGHỆ
OCR Engine	ABBYY OCR engine (proprietary)
Layout Analysis	Deep learning CNNs + NLP. Classify by appearance/pattern và text semantics
Table Extraction	Advanced table recognition với merged cells support
Data Validation	Multi-level validation: field-level, document-level, batch-level. Business rules engine
Output Format	XML, JSON, CSV, database export. REST API for cloud

---

6. NANONETS

STEP	CÔNG NGHỆ
OCR Engine	Proprietary AI OCR
Layout Analysis	Deep learning - template-free extraction. Learns from documents automatically
Table Extraction	AI extractors for line items, complex tables
Data Validation	Decision engines, human-in-the-loop, confidence scores
Output Format	JSON, XLS, CSV, XML. Direct integration với ERP/CRM

---

7. ROSSUM

STEP	CÔNG NGHỆ
OCR Engine	Proprietary transactional LLM (276 languages)
Layout Analysis	LLM-based understanding of document structure
Table Extraction	Extract line items, complex tables
Data Validation	Cross-validation với master data, ERPs, business rules. AI human collaboration
Output Format	JSON. Direct ERP integration (SAP, NetSuite, etc.)

---

8. MINDEE

STEP	CÔNG NGHỆ
OCR Engine	Proprietary AI OCR
Layout Analysis	Vision-aware pipeline, bounding boxes
Table Extraction	Line items, complex tables (merged cells). Vision models (không generic LLM)
Data Validation	Confidence scores, RAG for continuous learning, schema validation
Output Format	JSON, webhook support. Async processing cho large files

---

9. DOCPARSER

STEP	CÔNG NGHỆ
OCR Engine	Zonal OCR + Pattern recognition
Layout Analysis	Anchor keywords, zonal selection. Rule-based
Table Extraction	Column divider configuration
Data Validation	Rule-based extraction, confidence scores
Output Format	JSON, CSV, XML, Excel, Google Sheets. Zapier integration

---

10. PARSEUR

STEP	CÔNG NGHỆ
OCR Engine	AI document parser (template-free)
Layout Analysis	AI-based, learns from examples
Table Extraction	Line items capture, normalize data
Data Validation	Template-based, user feedback loop
Output Format	JSON, webhook, direct app integration

---

BẢNG SO SÁNH THEO TỪNG STEP

SOLUTION	OCR ENGINE	LAYOUT ANALYSIS	TABLE EXTRACTION	DATA VALIDATION	OUTPUT
Textract	Proprietary ML	Deep Learning	Basic merged cells	Confidence scores	JSON
Google Doc AI	Google Vision	DL + Enterprise OCR	Simple tables	Confidence	JSON
Azure	Azure AI Vision	Deep Learning	Complex tables	Accuracy scores	JSON
Adobe	Adobe Sensei	DL + Reading Order	Advanced	Confidence	JSON + CSV/XLSX
ABBYY	ABBYY OCR	CNN + NLP	Advanced + merged	Multi-level rules	XML/JSON/CSV
Nanonets	Proprietary AI	Template-free DL	Complex tables	Decision engine	JSON + ERP
Rossum	Transactional LLM	LLM-based	Line items	Cross-validation	JSON + ERP
Mindee	Proprietary AI	Vision-aware	Complex tables	RAG + confidence	JSON + Webhook
DocParser	Zonal OCR	Rule-based	Column config	Rule-based	JSON/CSV/XML
Parseur	AI Parser	AI Learning	Line items	Template-based	JSON + Webhook

---

KEY OBSERVATIONS

1. OCR Engine: Major players (AWS, Google, Azure, Adobe) sử dụng proprietary ML engines. Các startup (Rossum, Mindee, Nanonets) đầu tư vào LLM/vision models.

2. Layout Analysis: Xu hướng chuyển sang Deep Learning + NLP thay vì rule-based. ABBYY và Rossum tích hợp NLP capabilities.

3. Table Extraction: Đây là phần khó nhất. Adobe, ABBYY, Mindee claim support merged cells. Generic LLMs (như ChatGPT) thường "hallucinate" table structures - nên dùng vision-specific models.

4. Data Validation:

   • Big cloud providers: Confidence scores
   • Enterprise (ABBYY, Rossum): Business rules engine, multi-level validation
   • AI-first (Nanonets, Rossum): Human-in-the-loop, continuous learning

5. Output: JSON là chuẩn. ERP integration (SAP, NetSuite) là differentiation lớn cho Rossum, Nanonets. Async processing cần thiết cho files >10MB.

---

Research completed. Data gathered from official product documentation and websites.