How to Export Screenshot OCR to CSV with Coordinates and Confidence

Most OCR tools give you text. CaptureOCR gives you data: every row carries screen coordinates, line index, confidence score, OCR mode, processing time, and model hash. Ready for Excel, Python, or your data pipeline.

What’s Actually in the CSV

Export → CSV (UTF-8 with BOM) → open in Excel. Here’s every column:

Column Type Description
session_id UUID Unique session identifier
index int Display order (1, 2, 3… matches drawer)
x int Region left (physical pixels)
y int Region top (physical pixels)
w int Region width (physical pixels)
h int Region height (physical pixels)
line_index int Line number within region (0-based)
confidence float 0.0000–1.0000 (4 decimal places)
text string Recognized text (UTF-8)
mode string fast or accurate
elapsed_ms int OCR time for this region
model_hash string SHA-256 of model files (reproducibility)

Why Coordinates Matter

Use Case How Coordinates Help
Reproducibility Navigate to exact UI location for reproduction
DPI independence Physical pixels → works across 100%/125%/150% displays
Automation Feed coordinates to UI automation tools (pyautogui, etc.)
Audit trail Prove which screen area produced which data
Multi-monitor x,y are virtual desktop coordinates — knows which monitor

Physical pixels means coordinates are DPI-independent. On a 150% display, a region at logical (100, 100) with 150% DPR becomes physical (150, 150) — the CSV stores the physical values so they’re consistent regardless of display scaling.

Why Confidence Scores Matter

Confidence Range Meaning Action
0.99–1.00 Very high Trust completely
0.95–0.98 High Trust, quick visual check
0.90–0.94 Medium Review if critical
0.80–0.89 Low Likely needs correction
< 0.80 Very low Likely garbage, re-capture

Pro tip: In Excel, filter confidence < 0.90 → highlight rows → review only those.

Real CSV Sample (Actual Export)

session_id,index,x,y,w,h,line_index,confidence,text,mode,elapsed_ms,model_hash
a1b2c3d4-5678-90ab-cdef-1234567890ab,1,120,45,380,80,0,0.9984,Invoice #INV-2024-0315,fast,78,abc123def456...
a1b2c3d4-5678-90ab-cdef-1234567890ab,1,120,45,380,80,1,0.9921,Date: 2024-03-15,fast,78,abc123def456...
a1b2c3d4-5678-90ab-cdef-1234567890ab,2,120,140,380,60,0,0.9953,Vendor: Acme Corp,fast,65,abc123def456...

UTF-8 BOM — the file starts with \ufeff so Excel opens Chinese/Japanese correctly without “Data → Text to Columns.”

CSV vs Markdown vs JSON (.snipx)

Feature CSV Markdown JSON (.snipx)
Excel native ✅ ❌ ❌
Python/pandas ✅ ⚠️ ✅
Coordinates ✅ ✅ ✅ (normalized 0–1)
Confidence ✅ ✅ ✅
Model metadata ✅ ✅ ✅
Normalized boxes ❌ ❌ ✅ (0–1 relative)
Images included ❌ ❌ ✅ (PNG in ZIP)

Choose CSV for: Excel analysis, Python pandas, BI tools, data pipelines.
Choose .snipx for: Full reproducibility, sharing with team, re-export with different settings.

Python: Read CaptureOCR CSV

import pandas as pd

df = pd.read_csv('capture_export.csv', encoding='utf-8-sig')

# Filter low confidence
low_conf = df[df['confidence'] < 0.90]
print(f"Low confidence rows: {len(low_conf)}")

# Group by region
for region_idx, group in df.groupby('index'):
    print(f"Region {region_idx}: {len(group)} lines, avg conf {group['confidence'].mean():.4f}")

# Export low-confidence for review
low_conf[['index', 'line_index', 'confidence', 'text']].to_csv('review.csv', index=False)

Python: Reconstruct Regions from Coordinates

import pyautogui

# Click center of region 3
row = df[df['index'] == 3].iloc[0]
center_x = row['x'] + row['w'] // 2
center_y = row['y'] + row['h'] // 2
pyautogui.click(center_x, center_y)

Excel: Conditional Formatting for Confidence

  1. Select confidence column
  2. Home → Conditional Formatting → Color Scales
  3. Green (high) → Yellow (medium) → Red (low)
  4. Instant visual: green = trust, red = review

Data Pipeline Integration

Airflow/Dagster Task

def process_capture_csv(**context):
    df = pd.read_csv('/data/captures/latest.csv', encoding='utf-8-sig')
    
    # Quality gate
    assert df['confidence'].min() > 0.80, "Unacceptably low confidence"
    
    # Transform
    df['source'] = 'captureocr'
    df['ingested_at'] = pd.Timestamp.now()
    
    # Load to warehouse
    df.to_sql('ocr_captures', con=engine, if_exists='append', index=False)

dbt Model

-- models/staging/stg_captureocr.sql
SELECT
    session_id,
    index AS region_display_order,
    x, y, w, h,
    line_index,
    confidence,
    text,
    mode,
    elapsed_ms,
    model_hash,
    CURRENT_TIMESTAMP AS loaded_at
FROM 
WHERE confidence >= 0.80

Advanced: Session Package for Full Fidelity

Export .snipx → unzip → session.json contains normalized boxes (0–1 relative to region) — only in JSON, not CSV/Markdown. Use for precise overlay rendering.

{
  "id": "uuid",
  "mode": "fast",
  "selections": [
    {"index": 1, "rect_phys": [120,45,380,80], "screen_index": 0, "dpr": 1.5}
  ],
  "results": {
    "1": {
      "ok": true,
      "lines": [
        {"text": "Invoice #INV-2024-0315", "confidence": 0.9984, "box": [0.1, 0.1, 0.9, 0.2], "line_index": 0}
      ],
      "elapsed_ms": 78,
      "model_hash": "abc123..."
    }
  }
}

Compare: CaptureOCR CSV vs Other Tools

Feature CaptureOCR Windows Snipping Tool Snagit ABBYY
Coordinates ✅ Physical px ❌ ❌ ✅
Confidence ✅ 4 decimals ❌ ❌ ✅
Line index ✅ ❌ ❌ ✅
Model metadata ✅ ❌ ❌ ⚠️
UTF-8 BOM ✅ ✅ ✅ ✅
Batch export ✅ ❌ ❌ ✅

Automation Ideas

  1. Watch folder: New CSV → auto-ingest to database
  2. Quality gate: Fail CI if any confidence < 0.85
  3. Alerting: Low confidence → Slack notification
  4. Versioning: Store model_hash → detect model drift
  5. A/B testing: Compare fast vs accurate confidence distributions

Get Structured Data from Screenshots

  1. Install CaptureOCR ($2.99 one-time)
  2. Capture regions → batch OCR
  3. Export CSV → feed your pipeline

Get CaptureOCR from Microsoft Store

Related reads: