# Detect watermark region (first page, look for repeated gray text) first_page = doc[0] watermarks = [] for block in first_page.get_text("dict")["blocks"]: for line in block.get("lines", []): for span in line.get("spans", []): if span["color"] < 0.5: # dark gray/black threshold bbox = fitz.Rect(span["bbox"]) watermarks.append(bbox)
# Most watermarks are at same coordinates across pages common_rect = fitz.Rect() if watermarks: common_rect = watermarks[0] # simplify: take first pdf remove watermark github
This physically removes the text—even from copied text layer. Image watermarks (scan of a stamp, logo) require a different approach: # Detect watermark region (first page, look for
# Step 1: Generate a mask where watermark exists (manual ROI) convert input.pdf[0] -threshold 50% mask.png for i in $(seq 0 $(pdfinfo input.pdf | grep Pages | awk 'print $2')); do convert input.pdf[$i] mask.png -compose dst_out -composite page_$i.pdf done Step 3: Rebuild PDF and OCR pdfunite page_*.pdf no_watermark.pdf ocrmypdf no_watermark.pdf final_clean.pdf --deskew --clean | Tool | Stars | Method | Best
From a technical perspective, a watermark is just another layer of PDF content—text, vector art, or image—drawn over or under the main content. PDF’s stacking model makes removal possible via content filtering. | Tool | Stars | Method | Best for | |------|-------|--------|----------| | pdfrw + custom script | ~500 | Filter page contents by type | Text watermarks | | PyPDF2/PyMuPDF (fitz) | 6k+ | Remove annotations/overlay objects | Stamped watermarks | | pdfCropMargins | ~300 | Crop then scale | Edge watermarks | | OCRmyPDF + masking | 4k+ | OCR + regenerate | Image-based watermarks | | Stirling-PDF | 20k+ | GUI + CLI with “Remove Watermark” | Non-technical users |
And never remove watermarks to misrepresent ownership—that’s where engineering becomes forgery. This piece was assembled from real GitHub source analysis and PDF internals documentation. The code examples run on Python 3.8+ with PyMuPDF installed ( pip install PyMuPDF ).