pdfcpu2 is a fork of pdfcpu that adds offline OCR and text extraction while retaining the existing PDF processing commands. It builds as a single static binary, with no Python, cgo, or external services required.
pdfcpu2 ocr scan.pdf scan_ocr.pdf && pdfcpu2 extract -m text scan_ocr.pdf -This adds a searchable text layer to a scanned PDF and extracts the resulting text. The same pipeline can be used for RAG ingestion, agent tools, and batch document processing.
pdfcpu2 extract -m text: extract text from a PDF |
pdfcpu2 ocr: add an invisible, selectable text layer to scans |
|---|---|
 |
 |
The project wiki contains about 60 pages of documentation, including CLI usage, Go API examples, and getting-started guides.
No installation needed? A browser version is available at https://huzpsb.github.io/pdfcpu2/ — the same engine compiled to WebAssembly, processing everything locally in the page. See the Web UI wiki page for what it covers.
Horst Rutter, the maintainer of pdfcpu, is a member of the PDF Association. I admire his commitment to the PDF standard. But I genuinely need text extraction, OCR, and related features in my own work. Respect for his work does not remove that need, so I added them here and use them.
-
Single binary. The PP-OCRv6 detection and recognition models (about 6 MB) are embedded in the executable. OCR runs offline without API keys, tokens, rate limits, or data leaving the local machine. Windows, Linux, and macOS are supported on amd64 and arm64.
-
Also runs in the browser. The same engine compiled to WebAssembly powers a local-first web app at https://huzpsb.github.io/pdfcpu2/ — encrypt/decrypt, OCR, solidify, obfuscate, optimize, text extraction, watermark, split/merge, images→PDF, attachments, signatures, and batch processing run entirely in the page. No file is ever uploaded.
-
Pure Go OCR inference. The models run on a pure Go engine using fp32 scalar math and goroutine-based parallelism. No cgo, ONNX Runtime, or Python is required.
-
OCR benchmarks. PaddlePaddle's published benchmark reports the following results for PP-OCRv6 and several general-purpose VLMs:
Model Text detection Text recognition PP-OCRv6_medium 86.2 83.2 PP-OCRv6_small 84.1 81.3 PP-OCRv6_tiny (embedded) 80.6 73.5 Gemini-3.1-Pro 46.8 71.4 GPT-5.5 45.6 64.2 Qwen3-VL-235B 38.3 74.9 Source: PP-OCRv6 model card. The specialized models also have no per-page API cost.
-
PDFBox-compatible text extraction.
extract -m textfollows Apache PDFBox's default extractor. Behavior has been tuned through differential testing against PDFBox on a corpus of more than 1,200 real-world Chinese PDFs from different periods and generators. -
Page rendering. A self-contained pure-Go rasterizer (text including CJK/CID fonts, images, shadings, transparency groups) turns pages into PNG/JPEG images (
render).solidifyrebuilds pages as image-only pages to block text copying and content modification,obfuscategarbage fills the cmap/ToUnicode mappings of embedded CID fonts to block text extraction while pages keep rendering identically. Text drawn with fonts that are not embedded falls back to the bundled vivo Sans. -
Text replacement.
replaceremoves every occurrence of a find string from the content streams (page content and Form XObjects, whitespace-insensitive matching) and draws the replacement in place, anchored at the visually topmost-left matched glyph and preserving the original text matrix, font size, spacings, rise, render mode and fill color. Replacement glyphs come from one embedded vivo Sans subset that is reused and extended across runs, so replaced text stays selectable, extractable and copy-pastable; bold originals are approximated with fill+stroke. Output is optimized on write. -
CJK support. Text extraction handles ToUnicode CMaps (including Type42 and array-form
bfrange), embedded TrueType cmaps, legacy GBK/Big5, Adobe-GB1/CNS1 CID fallbacks, and a Latin-1 last resort. The OCR models are trained for Chinese and English. -
Repeatable OCR. Processed images are marked, so running
ocragain only processes new content. Text is added as an invisible layer mapped through each image's page transform, keeping selection aligned with the visible glyphs. -
CLI and API integration. Supports stdin/stdout streaming (
-), page selection, JSON output, andcontext.Contextcancellation in the Go API.
| Feature | Command | Status |
|---|---|---|
| Plain-text extraction (PDFBox-compatible) | pdfcpu2 extract -m text |
Available |
| Invisible OCR text layer for scanned PDFs | pdfcpu2 ocr |
Available |
| Page rendering / rasterization | pdfcpu2 render |
Available |
| Text replacement (embedded vivo Sans subsets) | pdfcpu2 replace |
Available |
| Content freezing (replace page content with its image) | pdfcpu2 solidify |
Available |
| Glyph obfuscation (garbage fill font cmaps and ToUnicode CMaps) | pdfcpu2 obfuscate |
Available |
# Text extraction
pdfcpu2 extract -m text in.pdf out.txt # whole document
pdfcpu2 extract -m text -p 1-3 in.pdf - # pages 1-3 to stdout
# OCR: make scans searchable and extractable
pdfcpu2 ocr scan.pdf scan_ocr.pdf # all pages
pdfcpu2 ocr scan.pdf scan_ocr.pdf -pages 1-10 # selected pages
# Rendering: one image file per page
pdfcpu2 render in.pdf outDir # PNG at 150 dpi
pdfcpu2 render -m jpg -q 90 --dpi 300 in.pdf outDir
# Solidify: rasterize pages and replace their content (anti copy/modify)
pdfcpu2 solidify in.pdf out.pdf # lossless PNG images
pdfcpu2 solidify -m jpg --dpi 200 in.pdf out.pdf
# Obfuscate: garbage fill the font mappings of embedded CID fonts
pdfcpu2 obfuscate in.pdf out.pdf # renders identically, extracts garbage
# Replace text: delete matches, redraw the replacement in place
pdfcpu2 replace -f "ACME" -r "CONTOSO" in.pdf out.pdf
pdfcpu2 replace -f "DRAFT" in.pdf out.pdf # empty replacement = delete onlyAs Go code:
conf := model.NewDefaultConfiguration()
ctx := context.Background()
n, err := api.OCRFile(ctx, "scan.pdf", "scan_ocr.pdf", nil, conf) // images processed
err = api.ExtractTextFile(ctx, "scan_ocr.pdf", "out.txt", nil, nil, conf)
// Or extract structured text with line coordinates:
texts, err := api.ExtractText(ctx, rs, []string{"1-3"}, nil, conf) // []pdfcpu.PageText
// Render pages to PNG files:
o := &render.Options{DPI: 150, Format: "png"}
err = api.RenderPagesFile(ctx, "in.pdf", "outDir", nil, o, conf)
// Or replace page content with its rendered image:
err = api.SolidifyFile(ctx, "in.pdf", "out.pdf", nil, o, conf)
// Or garbage fill the text extraction mappings of embedded CID fonts
// (renders identically, extracts garbage):
stats, err := api.ObfuscateFile(ctx, "in.pdf", "out.pdf", conf)
// Or replace text in place (occurrences replaced):
n, err = api.ReplaceTextFile(ctx, "in.pdf", "out.pdf", "ACME", "CONTOSO", nil, conf)
_ = stats // stats.FontFiles, stats.ToUnicode, stats.SkippedThe existing command set remains available, including validation, optimization, merge/split/trim, pages and boxes, stamps and watermarks, forms, encryption, permissions, signatures, attachments, PDF creation, n-up, booklets, and posters.
pdfcpu2 validate in.pdf
pdfcpu2 merge out.pdf a.pdf b.pdf
pdfcpu2 encrypt --opw owner --upw user in.pdf out.pdf # AES-256
pdfcpu2 stamp add "CONFIDENTIAL" "pos:tr, op:.6" in.pdf out.pdfgo install github.com/huzpsb/pdfcpu2/cmd/pdfcpu2@latest
# Or from a checkout:
go mod vendor && go build ./cmd/pdfcpu2As a library:
go get github.com/huzpsb/pdfcpu2pdfcpu2 is an independently maintained hard fork of pdfcpu, licensed under Apache 2.0. It is not guaranteed to track upstream releases. Issues and reasonable feature requests are welcome.
Existing behavior and API conventions remain compatible where practical. New code follows the existing package layering: model → pkg/pdfcpu → pkg/api → pkg/cli.
Upstream project: pdfcpu.
Apache License 2.0, same as upstream (see LICENSE.txt).
This software embeds vivo Sans Light as the fallback font used for rendering text whose fonts are not embedded in a PDF and as the glyph source for text replacement. vivo Sans is licensed by vivo Mobile Communication Co., Ltd. under the vivo Sans Font Knowledge Property License Agreement (see internal/fallbackfont/vivoSans-LICENSE.txt): the font must be credited inside this software, may not be modified or redistributed standalone, and the copyright notice and agreement are retained alongside the font. Using pdfcpu2 to render documents does not transfer any vivo Sans rights to the output images beyond what that agreement grants for works created with the font.
The embedded Adobe CID-to-Unicode tables are derived from Adobe's UCS2 CMap resources (Adobe BSD-style license; copyright notice retained in the generated file's header).