Although the Internet has given us a worldwide infrastructure on which to
build the universal library, much of the world knowledge, history, and
literature is still trapped on paper in the basements of the world's
traditional libraries. Many libraries and content owners are in the process of
digitizing their collections. While many such efforts involve the painstaking
process of converting paper documents to computer-friendly form, such as
based formats, the high cost of such conversions limits their
extent. Scanning documents, and distributing the resulting images
electronically is not only considerably cheaper, but also more faithful to the
original document because it preserves its visual aspect.
Despite the quickly improving speed of network connections and computers, the
number of scanned document images accessible on the Web today is relatively
small. There are several reasons for this.
The first reason is the relatively high cost of scanning anything else but
unbound sheets in black and white. This problem is slowly going away with the
appearance of fast and low-cost color scanners with sheet feeders.
The second reason is that long-established image compression standards and
file formats have proved inadequate for distributing scanned documents at high
resolution, particularly color documents. Not only are the file sizes and
download times impractical, the decoding and rendering times are also
prohibitive. A typical magazine page scanned in color at 100 dpi in
would typically occupy 100
, but the text would be hardly readable: insufficient for screen viewing and
totally unacceptable for printing. The same page at 300 dpi would have
sufficient quality for viewing and printing, but the file size would be 300
at best, which is impractical for remote access. Another major problem is that
a fully decoded 300 dpi color images of a letter-size page occupies 24
of memory and easily causes disk swapping.
The third reason is that digital documents are more than just a collection of
individual page images. Pages in a scanned documents have a natural serial
order. Special provision must be made to ensure that flipping pages be
instantaneous and effortless so as to maintain a good user experience. Even
more important, most existing document formats force users to download the
entire document first before displaying a chosen page. However, users often
want to jump to individual pages of the document without waiting for the
entire document to download. Efficient browsing requires efficient random
page access, fast sequential page flipping, and quick rendering. This can be
achieved with a combination of advanced compression, pre-fetching,
pre-decoding, caching, and progressive rendering. DjVu decomposes each page
into multiple components (text, backgrounds, images, libraries of common
shapes...) that may be shared by several pages and downloaded on demand. All
these requirements call for a very sophisticated but parsimonious control
mechanism to handle on-demand downloading, pre-fetching, decoding, caching,
and progressive rendering of the page images. What is being considered here
is not just a document image compression technique, but a whole platform for
DjVu is an image compression technique, a document format, and a software
platform for delivering documents images over the Internet that fulfills the
DJVU IMAGE COMPRESSION
The DjVu image compression is based on three technologies:
DjVuPhoto, also known as
is a wavelet-based continuous-tone image
compression technique with progressive decoding/rendering. It is best used
for encoding photographic images in colors or in shades of gray. Images are
typically half the size as
for the same distortion.
DjVuBitonal, also known as
is a bitonal image compression that takes
advantage of repetitions of nearly identical shapes on the page (such as
characters) to efficiently compress text images. It is best used to compress
black and white images representing text and simple drawings. A typical
300 dpi page in DjVuBitonal occupies 5 to 25
(3 to 8 times better than
DjVuDocument is a compression technique specifically designed for color
digital documents images containing both pictures and text, such as a page of
a magazine. DjVuDocument represents images into separately compressed layers.
The foreground layer is usually compressed with DjVu Bitonal and contains the
text and drawings. The background layer is usually compressed with DjVuPhoto
and contains the background texture and the pictures at lower resolution.
DJVU DOCUMENT DELIVERY PLATFORM
The DjVu technology is designed from the ground up to support the efficient
delivery of digital documents over the Internet. It provides various ways to
deal with multi-page documents, and various ways to enrich the content with
hyper-links, meta-data, searchable text, etc.
The DjVu format has an official MIME type of
which is the preferred content-type to be given by http servers for
DjVu files. Unofficial mime types used historically are
which may still be encountered. Ideally, clients should be configured
to handle all three. (For web server configuration help, see
Bundled multi-page documents
Bundled multi-page DjVu document uses a single file to represent the entire
document. This single file contains all the pages as well as ancillary
information (e.g. the page directory, data shared by several pages,
thumbnails, etc.). Using a single file format is very convenient for storing
documents or for sending email attachments.
When you type the
of a multi-page document, the DjVu browser plugin starts
downloading the whole file, but displays the first page as soon as it is
available. You can immediately navigate to other pages using the DjVu
toolbar. Suppose however that the document is stored on a remote web server.
You can easily access the first page and see that this is not the document you
wanted. Although you will never display the other pages the browser is
transferring data for these pages and is wasting the bandwidth of your server
(and the bandwidth of the Internet too). You could also see the summary of the
document on the first page and jump to page 100. But page 100 cannot be
displayed until data for pages 1 to 99 has been received. You may have to
wait for the transmission of unnecessary page data. This second problem (the
unnecessary wait) can be solved using the ``byte serving'' options of the
protocol. This option has to be supported by the web server, the
proxies, the caches and the browser. Byte serving however does not solve the
first problem (the waste of bandwidth).
Indirect multi-page documents
Indirect multi-page DjVu documents solve both problems. An indirect
multi-page DjVu document is composed of several files. The main file is named
the index file. You can browse a document using the
of the index file, just like you do with a bundled multi-page document. The
index file however is very small. It simply contains the document directory
of secondary files containing the page data. When you browse an indirect
multi-page document, the browser only accesses data for the pages you are
viewing. This can be done at a reasonable speed because the browser maintains
a cache of pages and sometimes pre-fetches a few pages ahead of the current
page. This model uses the web serving bandwidth much more effectively. It
also eliminates unnecessary delays when jumping ahead to pages located
anywhere in a long document.
Every DjVu image optionally includes so-called annotation chunks. The
annotation chunk is often used to define hyper-links to other document pages or
to arbitrary web pages. Annotation chunks can also be used for other purposes
such as setting the initial viewing mode of a page, defining highlighted zones, or
storing arbitrary meta-data about the page or the document.
Every DjVu image optionally includes a hidden text layer that associated
graphical features with the corresponding text. The hidden text layer is
usually generated by running an Optical Character Recognition software. This
textual information provides for indexing DjVu documents and copying/pasting
text from DjVu page images.
DjVu documents sometimes contain pre-computed page thumbnails.
DjVu documents sometimes contain a navigation chunk
containing an outline, that is, a hierarchical
table of contents with pointers to the corresponding
DJVUZONE AND DJVULIBRE
The DjVu technology was initially created by a few researchers in AT&T Labs
between 1995 and 1999. Lizardtech, Inc. (
) then obtained a commercial license from AT&T and continued
the development. They have now a variety of solutions for producing
and distributing documents using the DjVu technology.
The DjVuZone web site (
) is managed by the few AT&T Labs researchers who created the
DjVu technology in the first place. We promote the DjVu technology
by providing an independent source of information about DjVu.
Understanding how little room there is for a proprietary document format,
Lizardtech released the DjVu Reference Library under the
Public License in December 2000. This library entirely defines the
compression format and the elementary codecs. Six month later, Lizardtech
released an updated DjVu Reference Library as well as the source code of the
These two releases form the basis of our initial DjVuLibre software. We
modified the build system to comply with the expectations of the open source
community. Various bugs and portability issues have been fixed. We also
tried to make it simpler to use and install, while preserving the essential
structure of the Lizardtech releases.
The DjVuLibre software contains the following components:
A DjVuBitonal command line encoder. This soft-pattern-matching compressor
produces DjVuBitonal images from PBM images. It can encode images without loss,
or introduce small changes in order to improve the compression ratio. The
lossless encoding mode is competitive with that of the Lizardtech commercial
A DjVuDocument command line encoder for images with few colors. This encoder
is well suited to compressing images with a small number of distinct colors
(e.g. screen-shots). The dominant color is encoded by the background layer.
The other colors are encoded by the foreground layer.
Command line tools to edit DjVu metadata as XML files.
DJVU ENCODERS AND ANY2DJVU
DjVuLibre comes with a variety of specialized encoders,
for photographic images,
for bitonal images, and
for images with few distinct colors.
Although these encoders perform well in their specialized domain,
they cannot handle complex tasks involving segmentation and
Another solution is provided by the compression server at
This machine uses pre-lizardtech prototype encoders from AT&T Labs and
performs almost as well as the commercial Lizardtech encoders. Please note
that the Any2DjVu compression server comes with no guarantee, that
nothing is done to ensure that your documents will remain confidential, and
that there is only one computer working for the whole planet.
Numerous people have contributed to the DjVu source code during the
last five years. Please submit a sourceforge bug report to update the
Regis M. Chaplin,
Paul G. Howard,
Yann Le Cun,
Joseph M. Orost,