PDF Crop Studio

Crop images out of any PDF, by hand, with pixel precision.

Automatic PDF image extractors (XObject dumps, "smart" croppers, LLM vision tools) constantly get it wrong: they merge two pictures into one, cut off the bleed, grab a background rectangle instead of the product, or miss vector art entirely. PDF Crop Studio takes the opposite approach — it renders each page at high resolution and lets you drag the exact rectangle you want. What you select is what you get.

Built for catalogs, magazines, datasheets, zines, scanned docs — any PDF where you need clean image cutouts and the automatic tools keep failing you.

Open a PDF → Extract data → drag over a product → the label and price auto-fill from the page text → Save.

Features

• 📄 Open any PDF and flip through pages, rendered crisp at 3× zoom.
• 🔍 Zoom & pan with the mouse wheel to line up the perfect selection.
• 🖱️ Drag to crop — live pixel dimensions while you drag.
• 🪄 Auto-label from the PDF — hit Extract data and every crop's label (and price, when detected) is pre-filled from the text inside or just below your selection. No CSV, no retyping.
• 🎯 Key info only — toggle "Only items with a price" to keep real products/entries and drop greetings, headers and stray numbers.
• 🔢 Works with code-based catalogs — for catalogs whose items have a product code instead of a name (126031, W2155), the code becomes the entry and is detected on crop; no name required.
• 🔎 OCR fallback (Windows) — scanned or flattened PDFs with no text layer are read with the OCR built into Windows 10/11: whole pages during Extract data, and the exact crop region when you make a crop. Product codes (126031, W2155) are detected and pre-fill the ID field.
• 🗄️ Export to SQL — one click writes a ready-to-load .sql (id, label, price, category, page, image path) for all your crops.
• 🌐 English & Spanish — switch the whole UI language from the top-right; your choice is remembered.
• 🏷️ Label each crop with an optional ID, price and category.
• 📋 Optional CSV list — or load your own list of items (SKUs, figure numbers, names) to tag crops against and track what's still pending.
• 💾 Exports PNGs + manifest.json with page number and normalized bounding boxes, ready for any downstream pipeline.
• ⏯️ Resumable — close it and pick up exactly where you left off.
• 🌙 Clean dark UI, no cloud, no account, everything stays on your machine.

Install & run

Requires Python 3.9+.

git clone https://github.com/<your-user>/pdf-crop-studio.git
cd pdf-crop-studio
pip install -r requirements.txt
python run.py

Or open a PDF straight away:

python run.py path/to/catalog.pdf

Prefer a double-click, no-Python experience? See build_exe.md to build a standalone Windows .exe.

How it works

1. Open PDF (or pass one on the command line).
2. (Optional) Click Extract data — the app reads the PDF's text layer so crops can auto-fill their label and price from nearby text.
3. Navigate to a page; zoom in with the wheel until the image is comfortable.
4. Drag a rectangle over the region you want. Release to confirm.
5. The label (and price) come pre-filled from the text under your selection — just confirm or tweak, then hit Save.
6. The crop is written to <pdfname>_crops/crops/<key>.png and recorded in <pdfname>_crops/manifest.json.

Output

Everything lands in a folder named after your PDF, next to it:

mycatalog.pdf
mycatalog_crops/
├── project.json        # session state (so you can resume)
├── manifest.json       # flat list of all crops, for downstream tools
└── crops/
    ├── crop_0001.png
    ├── 10432.png       # named after the ID you typed, when you provide one
    └── ...

Each manifest.json entry looks like:

{
  "key": "10432",
  "id": "10432",
  "label": "Stainless steel thermos 1L",
  "category": "Kitchen",
  "price": 199.9,
  "page": 14,
  "bbox_norm": { "x0": 0.11, "y0": 0.32, "x1": 0.48, "y1": 0.71 },
  "bbox_px": { "x0": 594, "y0": 1728, "x1": 2592, "y1": 3834 },
  "width": 1998,
  "height": 2106,
  "file": "crops/10432.png",
  "source_pdf": "mycatalog.pdf"
}

bbox_norm is resolution-independent (0–1), so you can re-crop the same region from a higher-res render later if you need to.

Auto-labelling from the PDF text (Extract data)

Click Extract data (text + prices) after opening a PDF. The app scans the document's text layer, remembering where every line sits on the page. From then on, when you drag a crop:

• the label is pre-filled with the text that falls inside your selection (or the caption right below it, e.g. a product name under a photo);
• if a nearby line looks like a price ($, €, MXN, …), it's parsed and filled into the price field too.

The scan is cached next to your crops as extracted.json, so reopening the same PDF is instant. You can still edit any field before saving — nothing is forced.

What about scanned PDFs with no text layer?

Lots of catalogs are just images inside a PDF — there is no text to extract. On Windows 10/11, PDF Crop Studio falls back to the OCR engine built into the OS (installed with pip install winocr, no cloud involved):

• Extract data OCRs every page that has no text layer (you'll see the progress in the status bar).
• When you drag a crop on such a page, the app OCRs just your selection plus the caption area below it — small region, much higher accuracy — and pre-fills the label, price and even the product code into the ID field.
• Common OCR digit mistakes in prices are fixed automatically ($ I,839.OO → $1,839.00).

Heads-up: OCR quality depends on the source resolution. Crisp digital PDFs (≈150 DPI+) read great; a low-resolution rip (e.g. a 72 DPI catalog saved as a flat image) can only be read partially — the app tells you when it detects one, and you fix the odd wrong digit by hand after cropping.

Export to SQL

When you're done, click Export to SQL to write a single .sql file with one row per crop:

CREATE TABLE IF NOT EXISTS products (
  id TEXT, label TEXT, price REAL, category TEXT, page INTEGER, image TEXT
);
INSERT INTO products (id, label, price, category, page, image) VALUES
  ('10432', 'Mesa Roll', 1800, NULL, 1, 'crops/10432.png');

The image column points at the PNG in the crops/ folder next to the .sql, so the file loads straight into SQLite, Postgres or MySQL.

Language

Switch the interface between English and Español with the EN / ES buttons in the top-right corner. Your choice is saved for next time.

Using a CSV label list (optional)

Load a CSV to tag crops against a known list and track progress. The columns are auto-detected; a header like id,label,category works, and so does a plain two-column id,name file with no header.

id,label,category
10432,Stainless steel thermos 1L,Kitchen
10433,Glass food container set,Kitchen

With a list loaded, pick an item in the sidebar before you drag — its label is pre-filled, and it's marked ✓ once you've cropped it. "Only pending" hides the ones you've already done.

Mouse & keyboard

Two mouse modes, toggled with the Crop / Pan button in the sidebar:

• Crop mode (default): left-drag draws a crop rectangle.
• Pan mode: left-drag moves the page around.

In either mode you can also pan by dragging with the middle mouse button or by holding Space and dragging — handy for a quick nudge without leaving crop mode.

Shortcut	Action
Ctrl+O	Open a PDF
PageUp / PageDown	Previous / next page
Mouse wheel	Scroll up / down
Shift + wheel	Scroll left / right
Ctrl + wheel	Zoom at cursor
Ctrl + + / -	Zoom in / out
Middle-drag / Space+drag	Pan the page
Ctrl+F	Focus search
Enter (in search)	Jump to the highlighted item's page
Delete (in list)	Delete the selected crop
Esc	Cancel the current selection

Roadmap

Ideas for contributions (PRs welcome!):

• Export crops as JPG/WebP with quality settings
• Batch re-export at a chosen DPI from saved bbox_norm
• Rotate / straighten before saving
• macOS & Linux packaged builds
• Snap-to-content edge detection as an assist (never automatic)

License

MIT — see LICENSE. Do whatever you like with it.