discface

Transforms photos of CDs and DVDs into flat, top-down disc images. It removes perspective distortion, cuts out the disc with a transparent outer rim and centre hole, and rotates it so printed text reads upright. The disc boundary is detected geometrically, while text is read with doritex.

pip install discface

Example

import discface

disc = discface.extract("blank_dvd.jpg")       # photo of a blank DVD+R on a light table
print(disc.image.size, disc.image.mode)
print(f"turned by {disc.rotation:.1f}°, faint print enhanced: {disc.enhanced}")
disc.save("blank_dvd_disc.png")

Output:

(2048, 2048) RGBA
turned by -17.7°, faint print enhanced: True

blank_dvd.jpg

blank_dvd_disc.png

Perspective distortion is removed, leaving the disc neatly cut out with a transparent outer rim and centre hole. The image was rotated by −17.7° so that "Verbatim" reads horizontally. Because grey print on silver is often faint, it was also detected via edge-enhanced views (enhanced: True).

Taking the photo

discface is designed for discs placed on a plain, uniform background and photographed from roughly above. A sheet of white paper or a light grey tabletop works best, which is how all reference test photos were taken.

• Keep the entire disc in frame with some surrounding margin, ensuring the centre hole is visible.
• Shoot from roughly above. Moderate tilt is corrected automatically, and perspective is rectified precisely.
• Ensure even lighting. Moderate reflections are tolerated, and faint print benefits from enabling enhance.
• Capture a single disc per photo, keeping other circular objects out of frame.

In a benchmark where backgrounds around real discs were replaced while keeping the discs intact, rim detection rates were as follows:

On dark backgrounds, centre hole detection is less reliable (4/7). Patterned or cluttered backgrounds are not currently supported.

Command line

discface photos/ -o discs/

photos/IMG_4097.JPG -> discs/IMG_4097.png  turned 39.2°  59 words
photos/IMG_4100.JPG -> discs/IMG_4100.png  turned -1.7°  80 words
blank_dvd.jpg -> discs/blank_dvd.png  turned -17.7°  71 words  (faint print, edge-enhanced)

Inputs can be individual photos, directories, or a mix of both. Each disc is saved as <photo name><suffix>.<format> in the output folder.

Photos without a detectable disc are reported and skipped, returning exit code 1. You can also run the tool via python -m discface.

API

discface.extract(image, size=2048, margin=32, autorotate=True, enhance="auto", min_support=40, device="auto", detector=None) -> Disc Runs the three steps below in sequence. image can be a file path (with automatic EXIF orientation handling), a PIL.Image, or an RGB NumPy array. device selects the text recognition backend: "auto" uses an OpenCL GPU if present and falls back to CPU, while "cpu" forces CPU execution. Raises discface.DiscNotFound if no disc is detected.

The individual steps can also be called directly:

geometry = discface.locate("photo.jpg")                       # DiscGeometry: rim, hole, center in the photo
flat = discface.rectify("photo.jpg", geometry, size=2048)      # flat disc, RGBA, not turned
o = discface.orient(flat, enhance="always")                   # Orientation: rotation, support, turned, enhanced, words
upright = flat.rotate(o.rotation)

How upright orientation is decided

Every word detected on the disc votes for its reading direction. Large words carry more weight than fine print, and high-confidence detections outweigh uncertain ones. The dominant direction wins, so a few words printed sideways do not skew the result. Small text running circular along the rim is excluded, as it points in every direction. If too little text agrees (min_support), the disc retains its original orientation from the photo.

Faint print, such as grey text on a silver disc or lettering obscured by rainbow reflections, can also be read from two edge-enhanced views. These views emphasize character strokes, boost local contrast, and suppress smooth shading and reflections. The enhance option controls this behavior:

• auto (default): applies edge enhancement only when standard detection yields insufficient confidence. Discs that read well are processed without extra overhead.
• always: processes edge-enhanced views for every disc, adding roughly 1 to 2 seconds per image.
• off: never applies edge enhancement.

Performance

Benchmark times for a 24-megapixel photo after the initial warmup run (which compiles GPU kernels on first execution):

Both devices produce identical results. Only text recognition runs on the GPU; locating and rectifying the disc always run on the CPU.

License

MIT