Complex Table Recognition: How GLM-OCR Handles Merged Cells

Introduction

Table recognition is one of the most challenging tasks in document understanding. GLM-OCR achieves state-of-the-art performance in table recognition, handling complex structures that traditional OCR solutions struggle with.

Challenges in Table Recognition

Traditional OCR systems often fail with:

• Merged cells: Cells spanning multiple rows or columns
• Multi-level headers: Nested header structures
• Borderless tables: Tables without visible grid lines
• Complex layouts: Mixed text and numerical data
• Rotated tables: Tables at various angles

GLM-OCR Table Recognition Capabilities

GLM-OCR excels at all these scenarios thanks to its:

1. CogViT Visual Encoder: Understands spatial relationships
2. Deep Scene Optimization: Trained on real-world table data
3. Structured Output: Produces clean Markdown or JSON

Usage

Basic Table Recognition

# Using Ollama
ollama run glm-ocr Table Recognition: ./table.png

API Call

import requests

response = requests.post(
    "https://api.z.ai/api/paas/v4/layout_parsing",
    headers={"Authorization": "Bearer your-api-key"},
    json={
        "model": "glm-ocr",
        "file": "https://example.com/table.png"
    }
)

# Output is structured Markdown table
print(response.json()["content"])

Output Formats

Markdown Output

GLM-OCR outputs tables in clean Markdown format:

| Product | Q1 | Q2 | Q3 | Q4 | Total |
|---------|-----|-----|-----|-----|-------|
| Widget A | 100 | 150 | 200 | 180 | 630 |
| Widget B | 80 | 120 | 140 | 160 | 500 |
| **Total** | **180** | **270** | **340** | **340** | **1130** |

Handling Merged Cells

For tables with merged cells, GLM-OCR preserves the structure:

| Category | | Sales | |
|----------|----------|-------|-------|
| | Product | 2023 | 2024 |
| Electronics | Phone | 1000 | 1200 |
| Electronics | Laptop | 800 | 950 |
| Furniture | Chair | 500 | 600 |

Real-World Examples

Financial Reports

GLM-OCR handles complex financial tables with:

• Multiple header rows
• Currency formatting
• Percentage calculations
• Footnote references

Academic Papers

Scientific tables often include:

• Statistical data
• Confidence intervals
• P-values
• Multi-column layouts

Business Documents

Common business table types:

• Invoice line items
• Inventory lists
• Employee records
• Project timelines

Best Practices

Image Quality

• Resolution: Minimum 150 DPI recommended
• Contrast: Ensure clear distinction between text and background
• Orientation: Correct any rotation before processing

Complex Tables

For very complex tables:

1. Crop the table: Remove surrounding text
2. Enhance contrast: Improve visibility of grid lines
3. Split large tables: Process sections separately if needed

Post-Processing

import pandas as pd
from io import StringIO

# Parse Markdown table to DataFrame
markdown_table = """
| Product | Sales | Revenue |
|---------|-------|---------|
| A | 100 | $1000 |
| B | 200 | $2000 |
"""

# Convert to DataFrame for further processing
df = pd.read_csv(
    StringIO(markdown_table.replace("|", ",")),
    skipinitialspace=True
)

Performance Comparison

Troubleshooting

Common Issues

Issue: Columns misaligned Solution: Ensure the table image is not skewed

Issue: Missing cells Solution: Check image resolution and contrast

Issue: Merged cells not detected Solution: Use higher resolution image

Conclusion

GLM-OCR's table recognition capabilities make it ideal for processing complex business documents, financial reports, and academic papers. Its ability to handle merged cells, multi-level headers, and borderless tables sets it apart from traditional OCR solutions.

Related Articles

• GLM-OCR Quick Start Guide
• Formula Recognition with GLM-OCR
• Information Extraction Guide