Image to Text PDF#

Turn a finished raster image into a PDF while preserving the image exactly and adding a transparent text layer for selection, copying, and search.

Workflow#

1. Get the complete image.

• Treat the image as the visual source of truth.
• Keep the original image dimensions; the layout coordinates should use the same pixel coordinate space whenever possible.

1. Build a text-layer layout JSON.

• Prefer OCR or a vision model that returns text boxes over trying to infer positions manually.
• Use the user's source text to correct OCR transcription. OCR boxes determine position; the source text determines final copyable content.
• Read references/ocr-alignment.md when you need guidance for extracting, correcting, or prompting for box-level text.
• Read references/layout-json.md for the exact JSON schema.

1. Generate both PDFs.

bash

python scripts/compose_image_text_pdf.py \
  --image /path/to/image.png \
  --layout /path/to/layout.json \
  --output /path/to/image-text.pdf \
  --debug-output /path/to/image-text-check.pdf

For CJK or other non-Latin text, pass a Unicode font:

bash

python scripts/compose_image_text_pdf.py \
  --image /path/to/image.png \
  --layout /path/to/layout.json \
  --output /path/to/image-text.pdf \
  --debug-output /path/to/image-text-check.pdf \
  --font-file /path/to/NotoSansCJK-Regular.ttc

1. Inspect the debug PDF before delivering.

• The final PDF should look like the image-only source.
• The debug PDF should show highlighted text boxes and visible text where the hidden layer will be placed.
• If a highlighted box is shifted, fix the layout JSON, not the image.
• If copy/paste text is wrong, fix the text field in layout JSON, not the OCR image.

OCR Word Conversion#

When an OCR tool returns word-level boxes, convert them to line-level layout items:

bash

python scripts/ocr_words_to_layout.py \
  --ocr /path/to/ocr.json \
  --output /path/to/layout.json \
  --image-width 1536 \
  --image-height 2048 \
  --source-text /path/to/source.txt

The converter accepts common JSON shapes containing words, items, textAnnotations, or nested page/line/word objects. It groups nearby words into lines and can replace OCR text with the closest line from the source text when the match is strong.

Practical Rules#

• Keep text boxes line-level unless a paragraph must be selected as one unit. Line-level boxes are easier to position and debug.
• Do not try to match the rasterized font exactly. The hidden layer only needs close geometry and correct text.
• Use the visible inspection PDF as the validation artifact. It should make every embedded text span obvious.
• For generated images, ask the image-generation step to keep text in large, separated blocks. Dense tiny text is harder for OCR and manual correction.
• Preserve the image as the background instead of rebuilding the layout in presentation or web formats.