ocr

OCR

Use nong ocr as the only entrypoint for OCR and image-structure analysis. Do not route through Python OCR libraries or custom OCR wrappers.

Nong CLI Preflight

Read ../../.claude/references/nong-cli-preflight.md before the first Nong CLI command in a session. Confirm the nong CLI is installed and the ocr command group.

Modular: nong ocr routes to the standalone Angri450.Nong.Tool.Ocr dotnet tool. First use auto-installs. Command surface unchanged.

Prerequisites

Run once before OCR work:

nong ocr check-env --json

Cloud OCR and image-to-Word require this environment variable:

$env:PADDLEOCR_ACCESS_TOKEN = "<access-token>"

Get the cloud token from https://aistudio.baidu.com/account/accessToken. If the user has no token, use local OCR only after its smoke test passes.

Do not write real credentials into repository files, logs, or examples.

Route Table

User wants	Command	Reference
Environment check	nong ocr check-env --json	鈥?
Single-image text OCR	nong ocr local <img> --json	ocr-local.md
Batch directory OCR	nong ocr batch <dir> --pattern "*.png" --json	ocr-local.md
Video frame OCR + subtitles	nong ocr video <file> -o <dir> --json	鈥?
Screen region OCR (Win)	nong ocr screen --region x,y,w,h --json	鈥?
Camera capture OCR	nong ocr camera --device 0 --json	鈥?
Cloud OCR with layout	nong ocr cloud <input> -o <dir> --json	ocr-cloud.md
Image to Word	nong ocr to-word <input> -o <out.docx> --json	ocr-cloud.md
Image structure QA	nong ocr analyze-image <img> -o <dir> --json	image-analyzer.md
Model inventory	nong ocr models --json	鈥?
Install local model (v5)	nong ocr install-model pp-ocrv5-mobile --source <mirror> --json	runtime-chain.md
Install local model (v6 default)	nong ocr install-model pp-ocrv6-medium --json	runtime-chain.md
Install local model (v6 small)	nong ocr install-model pp-ocrv6-small --json	鈥?
Install local model (v6 tiny)	nong ocr install-model pp-ocrv6-tiny --json	鈥?

Implemented Commands

nong ocr check-env --json
nong ocr analyze-image <image.png> -o <analysis-dir> --json
nong ocr cloud <image-or.pdf> -o <ocr-out-dir> --json
nong ocr to-word <image-or.pdf> -o <out.docx> --json
nong ocr local <image.png> --json
nong ocr local <image.png> --force --json
nong ocr batch <dir> --pattern "*.png" --json
nong ocr batch <dir> --pattern "*.jpg" --recursive --json
nong ocr video <video.mp4> -o <dir> --fps 1 --json
nong ocr video <video.mp4> -o <dir> --dedup-threshold 8 --json
nong ocr screen --region 100,100,800,600 --json
nong ocr screen --json
nong ocr camera --device 0 --interval 2000 --count 5 --json
nong ocr models --json
# v5
nong ocr install-model pp-ocrv5-mobile --source https://mirrors.huaweicloud.com/repository/nuget/v3/index.json --json
nong ocr install-model pp-ocrv5-mobile --dry-run --json
# v6 (default: medium)
nong ocr install-model pp-ocrv6-medium --json
nong ocr install-model pp-ocrv6-small --json
nong ocr install-model pp-ocrv6-tiny --json
# pp-ocrv6 is an alias for pp-ocrv6-medium
nong ocr install-model pp-ocrv6 --json

Dispatch

1. For environment status, run nong ocr check-env --json.
2. For chart or diagram visual QA, run nong ocr analyze-image <image.png> -o <analysis-dir> --json.
3. For single-image text recognition, run nong ocr local <image.png> --json. Auto-selects v6 if installed, falls back to v5.
4. For scanning a directory of images, run nong ocr batch <dir> --pattern "*.png" --json.
5. For extracting text from video frames, run nong ocr video <video.mp4> -o <dir> --json. Outputs frames JSON + SRT subtitle file.
6. For OCR on a screen region (Windows), run nong ocr screen --region 100,100,800,600 --json.
7. For periodic camera capture OCR, run nong ocr camera --device 0 --interval 2000 --json.
8. For PDF, multi-page scans, table/layout reconstruction, or Word/NongMark annotation alignment, require PADDLEOCR_ACCESS_TOKEN, then run nong ocr cloud <input> -o <ocr-out-dir> --json.
9. For Word output from a scan or PDF, require PADDLEOCR_ACCESS_TOKEN, then run nong ocr to-word <input> -o <out.docx> --json.
10. For model inventory, run nong ocr models --json. Default model is pp-ocrv6-medium.
11. For v6 local deployment (recommended), run nong ocr install-model pp-ocrv6-medium --json.
12. For v5 local deployment (legacy), run nong ocr install-model pp-ocrv5-mobile --source https://mirrors.huaweicloud.com/repository/nuget/v3/index.json --json.

Boundaries

• ocr local, ocr batch: single-image text recognition. No PDF, no layout analysis, no table structure, no Word formatting.
• ocr video: requires opencv_videoio_ffmpeg*.dll in the native runtime. Samples frames (default 1 fps) with dHash deduplication. Outputs both JSON and SRT.
• ocr screen: Windows only. Uses Graphics.CopyFromScreen + GetSystemMetrics User32 P/Invoke.
• ocr camera: requires a connected camera and opencv_videoio_ffmpeg*.dll. Captures at configurable interval.
• ocr cloud and ocr to-word: require PADDLEOCR_ACCESS_TOKEN. Support PDF, multi-page, table/layout reconstruction.
• ocr analyze-image: structural image QA. Does NOT recognize text.
• If local OCR returns E005, do not suggest Python or pip. See runtime-chain.md.