arXiv Papers
G-Scholar Profile
CS Dept
DjVu Plugin for Win/Mac
DjVu Plugin for Unix
Websites that I maintain
NIPS Online
Yann's DjVu Page

image compression for digital libraries

DjVu (pronounced déjà vu) is an image compression technique, a file format, and a software platform, designed bring paper documents and high resolution photos to the Internet.

Most of the world's cultural and artistic production is still available only on paper. Distributing digitized versions of these documents on the Internet requires an image format that preserves the original aspect of the document, the readability of the text, and the quality of the pictures, while producing small file sizes even at high resolutions suitable for printing. Furthermore, even digitally produced documents often come in a variety of formats that are unsuitable for web distribution. While PDF partially addresses the problem of distributing digitally produced documents (although there are problems), it does not satisfactorily address the problem of distributing scanned documents (particularly color documents).

With DjVu, a black and white page scanned at 300 DPI occupies 5 to 30 KB, which is 3 to 8 times less than with the currently popular TIFF Group-IV format. A page from a magazine or an ancient document scanned at 300 DPI in color occupies 30 to 100KB in DjVu, while a JPEG version would be 5 to 10 times larger. A photograph or other continuous-tone image file occupies about half the size as a JPEG version with the same quality. Digitally produced documents are significantly smaller than their PDF equivalent, particularly if the document contains many pictures.

To compress black and white images, DjVu uses a technique called JB2 that attempts to find repeating shapes on the page (such as multiple occurences of a character). Coding the page comes down to coding representative prototypes of those shapes, together with a list of locations where they appear on the page.

To compress photos and continuous-tone images, DjVu uses a state-of-the-art method based on the mathematical concept of wavelets. This techniques, called IW44, is optimized for fast decoding and rendering. Its progressive decoding feature allow viewers to display an image very quickly, and then refine it as more bits arrive into the client machine.

For greyscale and color document, DjVu first separates the image into a background plane and a foreground plane. The foreground plane contains the text and line drawings, while the background plane contains the pictures and paper textures. The foreground plane is compressed with a variant of JB2, and the background with a variant of IW44, generally at lower resolution. With this technique, the text is kept at high resolution with very sharp edges, while the pictures are compressed at lower resolution.

DjVu was developed at AT&T Labs in Red Bank NJ by a research team composed of Yann LeCun, Leon Bottou, Patrick Haffner, Paul Howard, Pascal Vincent, Yoshua Bengio, and Bill Riemers. In April 2000, the DjVu technology was acquired by Seattle-based LizardTech Inc. which now commercializes DjVu.

An open source implementation maintained by Leon Bottou and friends is available here. It includes Unix versions of the browser plug-in, the viewer, decompressors, simple compressors, and various utilities.

Windows and Mac viewers and browser plug-ins are available free from LizardTech.

Free on-line conversion servers are available. Any2DjVu provides a simple way to upload documents and have them converted to DjVu while-u-wait. Bib2Web provides an automated way for researchers and students to produce publications page ready for posting on the web with full-text search capabilities.

Desktop compression software is available (free for non-commercial use) from LizardTech. Command-line compression tools are also available from LizardTech for Unix and Windows (with special conditions for educational and non-profit institutions).



Yann LeCun, Professor
The Courant Institute of Mathematical Sciences
Room 1221, 715 Broadway, New York, NY 10012, USA
tel: (212)998-3283

Copyright 2000-2004 Yann LeCun.

Yann LeCun, Le Cun, LeNet, DjVu, convolutional neural networks, machine learning, computer vision, pattern recognition, document imaging, image compression, digital libraries,