ocr-benchmark（OCR 基准测试）

OCR Benchmark v2.0.0

Multi-model OCR accuracy comparison with fuzzy line-level scoring, cost tracking, and PPT report generation.

Setup

1. Install dependencies

cd ~/.openclaw/workspace/skills/ocr-benchmark/ocr-benchmark
pip install -r requirements.txt

2. Configure environment variables

Set the variables for the providers you want to use:

# Bedrock (Claude models) — uses your existing AWS credentials
export AWS_REGION=us-west-2          # or your preferred region

# Gemini (Google AI Studio)
export GOOGLE_API_KEY=your_key_here

# PaddleOCR — OPTIONAL, skip if not available
export PADDLEOCR_ENDPOINT=https://your-paddle-endpoint
export PADDLEOCR_TOKEN=your_token    # optional auth token

Note on PaddleOCR: This provider requires an external API endpoint. If PADDLEOCR_ENDPOINT is not set, it is automatically skipped — no error. If you don't have a PaddleOCR endpoint, simply don't set the env var.

3. Prepare images

Place your images locally (.jpg, .png, .webp). There is no automatic image download — provide local file paths on the command line.

Quick Start

Run benchmark on images

python3 scripts/run_benchmark.py \
  --images img1.jpg img2.jpg img3.jpg \
  --output-dir ./results \
  --ground-truth ground_truth.json

Skip models with missing credentials (no error, just skips)

python3 scripts/run_benchmark.py \
  --images img1.jpg \
  --auto-skip \
  --output-dir ./results

Run only specific models

python3 scripts/run_benchmark.py \
  --images img1.jpg \
  --models opus sonnet gemini3pro \
  --output-dir ./results \
  --ground-truth ground_truth.json

Score-only mode (re-score without re-running OCR)

python3 scripts/run_benchmark.py \
  --score-only \
  --output-dir ./results \
  --ground-truth ground_truth.json

Generate PPT report from scored results

python3 scripts/make_report.py \
  --results-dir ./results \
  --images img1.jpg img2.jpg img3.jpg \
  --scores ./results/scores.json \
  --output report.pptx

Workflow

1. Prepare images — collect your .jpg / .png files locally
2. Run benchmark — run_benchmark.py calls each model, saves {image}.{model}.json
3. Create ground truth — see references/ground-truth-format.md for format
4. Score — run with --ground-truth to produce scores.json and a terminal table
5. Report — make_report.py generates a shareable .pptx

Environment Variables

Missing variables: If a model's required env var is missing, it is automatically skipped with a warning. Use --auto-skip for completely silent skipping.

Available Models

See references/models.md for full model IDs, pricing, and provider notes.

Scoring Logic (v2)

Scoring uses fuzzy line-level matching with Levenshtein edit distance (pure Python stdlib, no extra dependencies).

For each ground truth line, the best-matching model output line is found and classified:

Additionally, EXTRA lines are detected: model output lines that don't correspond to any ground truth line.

Normalization strips: whitespace, apostrophes/quotes (', ', ` `), common punctuation (*, ✓, ，, 、, ：, （）, 【】` etc.), then lowercases.

Example terminal output

========================================================================
  OCR BENCHMARK RESULTS
========================================================================
  #    Model                        Score  Details
------------------------------------------------------------------------
  🥇   Gemini 3.1 Pro               98.7%  Image001: 99% | Image002: 98%
  🥈   Claude Opus 4.6              88.3%  Image001: 90% | Image002: 87%
  🥉   Claude Sonnet 4.6            85.1%  Image001: 86% | Image002: 84%
  4.   Gemini 3.1 Flash-Lite        82.0%  ...
========================================================================

  📄 Image001
  ──────────────────────────────────────────────────────────────────────
  ┌─ Claude Opus 4.6 (90.0%)
  │  ✅ EXACT   │ 小胡鸭
  │  🟡 CLOSE   │ GT: Sam's Coffee
  │             │ Got: Sams Coffee  [dist=2]
  │  🟠 PARTIAL │ GT: 浓郁香气
  │             │ Got: 浓都香气  [dist=1]
  │  ❌ MISS    │ GT: 净含量580克
  │  ⚠️  EXTRA lines (1):
  │     + "Product of China"
  └──────────────────────────────────────────────────────────────────────

Output Files

Each OCR run produces {image}.{model}.json:

{
  "text_extracted": ["line1", "line2", ...],
  "brand": "...",
  "product_name": "...",
  "net_weight": "...",
  "ingredients": ["..."],
  "other_fields": {},
  "model": "Claude Opus 4.6",
  "model_key": "opus",
  "latency_seconds": 23.5,
  "input_tokens": 800,
  "output_tokens": 500
}

Scoring produces scores.json with per-image, per-line, per-model results.

Key Findings (2026-03, product packaging)

Human-verified ranking:

• Gemini 3.1 Pro (98.7%) — Best accuracy, ~$0.006/image
• Claude Opus 4.6 (92.3%) — High accuracy; occasional missed details
• Gemini 3.1 Flash (89.7%) — Best speed/cost ratio, 9.7s
• Claude Sonnet 4.6 (88.5%) — Stable structured output
• PaddleOCR (67.9%) — Free, character errors on packaging
• Claude Haiku 4.5 (42.3%) — Poor Chinese OCR

Lesson: Never assume any model is ground truth. Human verification is essential.