Ffmpeg Ocr Logo

Ffmpeg Ocr Logo

Context: $ARGUMENTS

OCR, logo removal, and QR operations inside ffmpeg are filter-graph operations. They run on the decoded frame, emit results as frame metadata under lavfi.<filter>.*, and/or modify pixels in place. All are re-encode only — you cannot stream-copy through any of these filters.

Several of the filters in this skill are build-flag gated — ocr needs --enable-libtesseract, quirc needs --enable-libquirc, and qrencode needs --enable-libqrencode. Many static / minimal ffmpeg builds ship without these. Always check the build first (Step 1) before writing a pipeline.

Quick start

• Read text from video: → Step 1 (check ocr) → Step 2 (pick "OCR") → Step 3
• Hide a known rectangle (e.g. bug/logo at fixed coords): → Step 2 (pick "delogo") → Step 3
• Remove a logo with an alpha mask PNG: → Step 2 (pick "removelogo") → Step 3
• Find and hide a recurring graphic (reference image): → Step 2 (pick "find-rect + cover-rect") → Step 3
• Decode QR codes from frames: → Step 1 (check quirc) → Step 2 (pick "QR decode") → Step 3
• Burn a QR code overlay into video: → Step 1 (check qrencode) → Step 2 (pick "QR encode") → Step 3

When to use

• Extracting burned-in text (lower-thirds, scoreboard tickers, timestamps) live from a stream.
• Removing a broadcaster bug / channel logo from a captured feed.
• Blurring license plates or faces at a consistent screen location.
• Embedding / reading QR watermarks for pipeline tracing or ad-insertion cues.

Not this skill: Pre-recorded high-accuracy OCR — extract frames with ffmpeg-frames-images then run tesseract directly for better accuracy. True object detection / face detection — use an external ML model. Arbitrary watermark removal without a known shape — ffmpeg has no generic inpainting.

Step 1 — Check your ffmpeg build

Run once before building any pipeline:

ffmpeg -hide_banner -filters 2>/dev/null | grep -E "ocr|find_rect|cover_rect|delogo|removelogo|qrencode|quirc"

Or use the helper:

uv run ${CLAUDE_SKILL_DIR}/scripts/ocrlogo.py check-build

Expected results:

If a filter is missing, either rebuild ffmpeg with the flag or fall back to the OS tool (tesseract, zbarimg, qrencode CLI) on extracted frames.

Step 2 — Pick the operation

A. OCR — read text from frames

ffmpeg -i in.mp4 \
  -vf "ocr=datapath=/usr/local/share/tessdata/:language=eng,metadata=mode=print:file=ocr.txt" \
  -an -f null -

• datapath is the Tesseract tessdata directory (contains eng.traineddata etc.). If omitted, ffmpeg uses Tesseract's compile-time default.
• metadata=mode=print:file=ocr.txt dumps lavfi.ocr.text for every frame. Without it, results stay in memory and are lost.
• Add -an to skip audio, -f null - to discard video output.
• Whitelist / blacklist characters via whitelist=0123456789: (useful for clocks) and blacklist=....

OCR a single frame at 10s:

ffmpeg -ss 10 -i in.mp4 -vframes 1 \
  -vf "ocr=datapath=/usr/local/share/tessdata/,metadata=mode=print" \
  -f null - 2>&1 | grep "lavfi.ocr.text"

B. delogo — erase a known rectangle

ffmpeg -i in.mp4 \
  -vf "delogo=x=10:y=10:w=200:h=80:show=0" \
  -c:v libx264 -crf 20 -c:a copy out.mp4

• x,y is the top-left of the rectangle; w,h its size.
• show=1 draws a green outline — use when tuning coordinates, then set back to 0.
• The rectangle is filled by interpolating the 1-pixel ring just outside it. Works well on solid colors, poorly on textured or semi-transparent logos.
• If edges bleed, extend the clearing zone via band=2 (default 1).

C. removelogo — erase via alpha-mask PNG

ffmpeg -i in.mp4 \
  -vf "removelogo=logo_mask.png" \
  -c:v libx264 -crf 20 -c:a copy out.mp4

• The mask must be the same resolution as the video.
• White (255) = logo pixel to remove; Black (0) = keep untouched. This is opposite of the typical alpha convention.
• Build the mask by screen-grabbing a black frame showing the logo, thresholding in GIMP/ImageMagick, then eroding once or twice.

D. find-rect + cover-rect — find a reference image and hide it

# ref.pgm must be gray8 (convert via ImageMagick: magick ref.png -colorspace gray ref.pgm)
ffmpeg -i in.mp4 \
  -vf "find_rect=object=ref.pgm:threshold=0.3,cover_rect=cover=blank.jpg:mode=cover" \
  -c:v libx264 -crf 20 -c:a copy out.mp4

• find_rect scans each frame for the reference sub-image via cross-correlation. The matched rectangle's position is written to lavfi.rect.{x,y,w,h,score}.
• cover_rect reads those metadata entries from its upstream filter — they must be in the same filter chain.
• cover= image must be yuv420 (e.g. a plain .jpg). Use mode=blur for a soft blur instead of an image patch:

find_rect=object=ref.pgm:threshold=0.3,cover_rect=mode=blur

• Tune threshold (0.01 = exact match, 0.99 = almost anything). Start at 0.3–0.5.
• Restrict search area with xmin/ymin/xmax/ymax for speed.
• discard=1 drops frames where no match was found (useful when you only want the detection hits).

E. QR decode (quirc)

ffmpeg -i in.mp4 \
  -vf "quirc,metadata=mode=print" \
  -an -f null - 2>&1 | grep "lavfi.quirc"

Emitted metadata per frame with a detected QR:

• lavfi.quirc.count=N (only set if ≥ 1 found)
• lavfi.quirc.0.payload=<decoded text>
• lavfi.quirc.0.corner.0.x … lavfi.quirc.0.corner.3.y (four corner coords)

F. QR encode (qrencode) — overlay QR on video

ffmpeg -i in.mp4 \
  -vf "qrencode=text='https://example.com':q=w*0.25:x=w-q-20:y=20:level=H" \
  -c:v libx264 -crf 20 -c:a copy out.mp4

• text or textfile supplies the payload.
• qrcode_width / q is the rendered size (accepts expressions referencing w, h).
• x, y position the top-left corner — same expression DSL as overlay/drawtext.
• level is L/M/Q/H (H = most redundancy, larger code).
• foreground_color / background_color take any ffmpeg color name.

Step 3 — Run the command and parse the metadata

When you use metadata=mode=print, ffmpeg emits a block per frame to stdout (or the file set by file=):

frame:42   pts:42000    pts_time:1.68
lavfi.ocr.text=Hello World
lavfi.ocr.confidence=87 91 93

Parse with regex:

import re
OCR_RE  = re.compile(r"^lavfi\.ocr\.text=(.*)$", re.M)
QR_RE   = re.compile(r"^lavfi\.quirc\.(\d+)\.payload=(.*)$", re.M)
RECT_RE = re.compile(r"^lavfi\.rect\.(x|y|w|h|score)=([\d.]+)$", re.M)

Or use the helper script:

uv run ${CLAUDE_SKILL_DIR}/scripts/ocrlogo.py ocr --input in.mp4 --output ocr.txt
uv run ${CLAUDE_SKILL_DIR}/scripts/ocrlogo.py qr-decode --input in.mp4
uv run ${CLAUDE_SKILL_DIR}/scripts/ocrlogo.py find-cover --input in.mp4 --output o.mp4 \
     --reference logo.pgm --cover blur

Available scripts

• scripts/ocrlogo.py — subcommands: check-build, ocr, delogo, removelogo, find-cover, qr-decode, qr-encode. Stdlib only. Supports --dry-run and --verbose.

Reference docs

• Read references/filters.md for full option tables, per-OS tessdata paths, regex parsing patterns, a logo-removal decision tree, and recipe book (plate blur, live-TV logo removal, QR watermark).

Gotchas

• ocr needs --enable-libtesseract. Static / minimal ffmpeg builds (including several Homebrew bottles and the official ffmpeg.org static releases) don't have it. Check with ffmpeg -filters | grep ocr first. On macOS use brew install ffmpeg (full, not ffmpeg@...-slim); on Linux check your distro package or build from source.
• datapath is the tessdata directory, not a file. It points at a folder that contains <lang>.traineddata. Defaults:
  • Linux (apt): /usr/share/tessdata/ or /usr/share/tesseract-ocr/<ver>/tessdata/
  • macOS Homebrew: /usr/local/share/tessdata/ (Intel) or /opt/homebrew/share/tessdata/ (Apple Silicon)
  • Windows: C:/Program Files/Tesseract-OCR/tessdata/
• Language files must match the language option. language=deu requires deu.traineddata in the datapath. Install extras via brew install tesseract-lang / apt install tesseract-ocr-deu.
• OCR accuracy drops hard on compressed video. Sample at high bitrate, pre-upscale (scale=iw*2:ih*2), or pre-threshold (eq=contrast=1.5,curves=preset=increase_contrast). For production OCR on a library, extract frames (ffmpeg-frames-images) and run tesseract directly — far more accurate than ffmpeg's in-graph ocr filter.
• ocr is best for LIVE or streaming use cases where you want text as frame metadata in a single pass (cropdetect-style). Don't use it for one-off transcription.
• metadata=mode=print format prints a frame:N pts:M pts_time:T header then lavfi.ocr.text="..." lines. Multi-line OCR output gets escaped with \n. Parse with a state machine or the regex pattern in Step 3.
• delogo works best on solid-color bugs (station logos, simple graphics). Struggles on textured, semi-transparent, or animated logos — the interpolation visibly smears.
• removelogo mask convention is inverted from intuition — white pixels mean "this is the logo, remove it," black means "keep." The mask must be the exact resolution of the video.
• find_rect needs a gray8 reference (.pgm or 8-bit grayscale .png). Convert with magick ref.png -colorspace gray ref.pgm or ffmpeg -i ref.png -pix_fmt gray -y ref.pgm.
• cover_rect's cover image must be yuv420. Use .jpg or convert via ffmpeg -i cover.png -pix_fmt yuv420p cover.jpg.
• find_rect + cover_rect must share a filter chain. They communicate via frame metadata (lavfi.rect.*); a -filter_complex with separate branches breaks the hand-off. Chain them with a comma.
• find_rect finds at most ONE instance per frame. Multiple logos of the same shape on one frame need multiple passes or a different approach.
• threshold semantics are inverted from what you'd guess — 0.01 means "only exact matches" (strict), 0.99 means "almost anything matches" (loose). Start at 0.3.
• quirc / qrencode are rare — both require explicit build flags (--enable-libquirc, --enable-libqrencode) and only landed in FFmpeg 7.1. On older or slimmer builds, decode with zbarimg and encode with the qrencode CLI on extracted frames.
• QR metadata key misspelling: the official docs show lavfi.quirc.N.coreer.M.y (sic — this is actually corner in the source). Match both spellings defensively if parsing.
• qrencode defaults to black-on-white and top-left placement. For watermarking a bright scene, swap to foreground_color=white:background_color=black and scale down to 10–15% frame width.
• All filters in this skill require a re-encode — you cannot -c copy through them. Pair with a reasonable -crf / -preset in the encoder call.

Examples

Example 1 — OCR a news ticker every second

ffmpeg -i news.mp4 \
  -vf "fps=1,ocr=datapath=/usr/local/share/tessdata/:language=eng,metadata=mode=print:file=ticker.txt" \
  -an -f null -
grep "^lavfi.ocr.text=" ticker.txt

fps=1 samples 1 frame per second before OCR to cut cost by 24–30x.

Example 2 — Remove a broadcaster bug in the top-right

First locate the bug coordinates with show=1:

ffplay -vf "delogo=x=1720:y=40:w=180:h=80:show=1" live.mp4

Then render the cleaned file:

ffmpeg -i live.mp4 \
  -vf "delogo=x=1720:y=40:w=180:h=80" \
  -c:v libx264 -crf 20 -c:a copy clean.mp4

Example 3 — Auto-blur a recurring reference object

Wherever the reference image appears, blur it.

# Prepare gray8 reference
magick ref.png -colorspace gray ref.pgm

# Auto-blur
ffmpeg -i surveil.mp4 \
  -vf "find_rect=object=ref.pgm:threshold=0.4,cover_rect=mode=blur" \
  -c:v libx264 -crf 20 -c:a copy blurred.mp4

Example 4 — Decode QR codes once per second

ffmpeg -i tracer.mp4 -vf "fps=1,quirc,metadata=mode=print:file=qr.log" -an -f null -
grep "payload=" qr.log

Example 5 — Embed a tracking QR code bottom-right

ffmpeg -i in.mp4 \
  -vf "qrencode=text='job-id=12345':q=128:x=w-q-20:y=h-q-20:level=Q:\
foreground_color=black:background_color=white@0.85" \
  -c:v libx264 -crf 20 -c:a copy out.mp4

Troubleshooting

Error: No such filter: 'ocr'

Cause: Your ffmpeg build doesn't include libtesseract. Solution: Check with ffmpeg -filters | grep ocr. Reinstall (brew reinstall ffmpeg — Homebrew's default is full-featured) or build from source with --enable-libtesseract. Alternatively, extract frames with ffmpeg-frames-images and run tesseract directly.

Error: Failed to init ocr (Failed loading language 'eng')

Cause: Tessdata missing or datapath wrong. Solution: Verify the folder exists and contains eng.traineddata: ls /usr/local/share/tessdata/. On macOS also install brew install tesseract. Pass the correct path via datapath=.

OCR output is gibberish / empty

Cause: Source video too compressed, text too small, or font antialiased too aggressively. Solution: Sample at higher resolution, pre-scale (scale=iw*2:ih*2:flags=lanczos), pre-contrast (eq=contrast=1.4), or crop to the text region first (crop=W:H:X:Y). For best accuracy, extract frames and run tesseract standalone.

delogo leaves a visible smear

Cause: Logo extends beyond the specified rectangle, or is textured. Solution: Widen w/h by 4–8 pixels, then set band=2. For textured logos try removelogo with a mask PNG instead.

removelogo: Mask file does not match video size

Cause: Mask PNG resolution differs from the video stream resolution. Solution: Resize the mask to match: magick mask.png -resize 1920x1080! mask_fixed.png. Must be an exact match.

find_rect never matches

Cause: Reference image not in gray8, or threshold too strict. Solution: Confirm gray8: ffprobe -show_streams ref.pgm | grep pix_fmt should say gray. Relax threshold (0.5 → 0.7). Reduce the reference to a tight crop of the logo — extra border content kills correlation score.

No such filter: 'quirc' / 'qrencode'

Cause: ffmpeg built without libquirc / libqrencode. Solution: Confirm with ffmpeg -filters | grep -E 'quirc|qrencode'. Either rebuild with --enable-libquirc --enable-libqrencode (FFmpeg 7.1+ only), or fall back to zbarimg (decode) / qrencode CLI (encode) on extracted frames.