Digital Access to Rare Documents
New Image Formats and Approaches for Document Delivery and Their Comparison with
Adolf Knoll, National Library of the Czech Republic
Efficient delivery of image data consists of an acceptable reduction of information. The paper concerns efficiency of new mixed
raster content approaches in comparison with traditional methods such as changes in resolution, colour depth, or compression. The
paper is complemented with many comparison tables and applicable samples
To be able to read this study, you should have a browser, which supports
the PNG image format. The latest versions of MSIE and Netscape do this.
If you want to see also special samples, it will be necessary for you to
install also two plug-ins for viewing the LuraDocument and the DjVu images.
Install the newest versions of plug-ins - in case you have older versions,
please de-install them prior to install the new ones. You will find the
plug-ins on the web sites of the Luratech
Company and the Lizardtech
Inc. They will be referred as browser plug-ins.
I Traditional methods
I.1 Reduction of space resolution
I.2 Reduction of brightness
I.4 Best lossless and lossy
standards and their emerging follow-on
II Mixed raster content
II.1 How to segment the image?
II.2 Testing a mixed raster
content page delivery
LDF with DjVu
III Wider comparisons with traditional approaches
III.1 True colour - 24-bit image-
16,7 million colours
IV Further reduction
of colour depth
V Textual document in shades of grey
VII Enclosed samples
Delivery of scanned documents in image format over networks requires
transfer of large quantity of data. In order to reduce it and make the
life of end users easier, we should reduce the volume of image files. We
have several possibilities how to do it, but in practice the best solutions
consist in combining two or more of them.
Here are the methods:
Reduction of space
The space (spatial) resolution is most frequently called simply resolution
and it says how many dots, i.e. elements of the picture (picture elements
= pixels) are used to express information found on a unit of length. The
most used unit of length is inch; therefore, the resolution is given in
per inch (dpi). However, it can be measured also per other units of
length as, for example, meters or centimetres.
The usual desktop scanners scan by default true colour images (frequently
defined as 16 million colours) at 100 dpi and bitonal images at
300 dpi that is good enough for OCR. If we have an image of 10 x 8 inches, then
we need 1000 x 800 dots to express it digitally so that we use 800,000
dots or pixels for the whole image. For 16 million colours we need 3 bytes
per pixel; therefore, our scanned image will have 3 x 800,000 bytes = 2,400,000
bytes = 2.4 MB.
Now if we reduce the space resolution from 100 dpi to 80 dpi, which
may be quite affordable for some types of images, we will need for the
same image 800 x 640 pixels, i.e. 3 x 512,000 bytes = 1.536 MB. Thus it
looks that by mere reduction of resolution we have saved 864 KB that is
a very substantial saving, but, of course, not everywhere this can be applied
in order not to distort the original image in an undesirable
If we express in other words what we have done, we see that for the
same information we used only 80% of resources to express it.
If the original information had fine details, lower quantity of picture
elements (pixels) to represent them means reduction of the quality of recorded
information. In other words, we compressed reality more with 80 dpi than
we did it with 100 dpi. This fact can be easily demonstrated when we enhance
a part of the scanned image to look closer at its details.
In the above image, the left side is the engine scanned at 100 dpi,
while the right image shows the same engine scanned at 80 dpi that seems
to give more contrast.
Here we can see a detail image of a part of the carburettor of the same
engine: left side at 100 dpi and right side at 80 dpi. We can discern that
thanks to higher space resolution the left image is smoother than the right
one, while the right image seems to have more contrast. We can use these
characteristics successfully in our work.
Reduction of brightness
The brightness resolution is more frequently called colour depth
of the image. It says how many colours are used to express one pixel.
If we imagine one pixel of the image as a square and if we use for this
square only 1 bit of information, this square can have only two values,
i.e. 0 or 1 that means black or white colours.
However, if we want to reproduce colour, we must use more information
for one pixel: for example, 4 bits, 8 bits (1 byte), or 24 bits (3 bytes);
accordingly, we have 4-bit, 8-bit, or 24-bit images. If we count colours
in place of pixels, we obtain the following numbers of colours:
|Number of bits per 1 pixel
||Number of colours
black-and-white, bitonal, bi-level
To express reality in a true way we use 24 bits per pixel that gives
us more than 16 millions of colours. Per one pixel we need then 3 bytes
of information and - as counted more above - for our image it takes 2,4
MB. If we dare to reduce the volume of information per one pixel from 24
bits to 8 bits, we need then 3 times less storage space for our image,
i.e. 800 KB. Thus we may save substantial space, but we must test, whether
we can afford such a substantial loss of information, because all these
of over 16 million colours are recalculated into only 256 colours in a
palette that is individual for our image.
This can be easily detected on the below photograph that was scanned
at 16 million colours. The number of colours was reduced onto 256 and 16
to demonstrate the loss of information. Here we have the original true
The second image has the reduced number of colours onto 256. The reduction
was optimized to minimize the loss of information (error diffusion method).
We can hardly observe the difference, but if we enhance the same detail
from the two images and compare the pixels from the same areas, we will
see how the 16 million colours have been recalculated into a palette of
256 colours. This happens, for example, automatically during the conversion
of a true colour image (uncompressed BMP or TIFF, TIFF/LZW, JPEG, TGA…)
In case that we reduce the number of colours onto only 16, the difference
is evident in spite of our trying to minimize the losses of information
by application of an appropriate error diffusion method. We can examine
pixel by pixel on the same positions to see a rough reduction of information:
We may also scan images in shades of grey. If applying this to the images
of engines from above the result is as follows: 256 shades of grey at 100
dpi (left) and 80 dpi (right).
In some cases - simple colour objects - we may reduce the number of
colours - shades of grey - to only 16 and thus to reduce the volume of
storage space needed onto 400 KB (in our sample if applicable). However, in most cases, the loss of information
is so high that we do not do it.
Here we have the same image in 16 shades of grey at 100 dpi (left) and
80 dpi (right). We can observe the reduction of detail, but the images
are still usable.
We can even try to optimize the image for 1-bit using error diffusion
methods such as Floyd-Steinberg, Stucki, or Burkes. For some type of usage
it can be still applicable, but the reduction of colours onto black-and-white
image can have even worse results if we apply an inappropriate colour reduction
method. Such an image can be hardly used, but sometimes the choice is very
difficult, because the method, which is good for image (error diffusion)
to produce the so-called halftones, is very bad for text. On the contrary,
the colour reduction method good enough for text (nearest colour reduction)
is inappropriate for image.
This is a result after application of the Floyd-Steinberg colour
This is a result after application of the nearest colour reduction
If there is also text with images, we need to apply rather the latter
method to be able to enable OCR or to reduce the volume of data files especially
when applying lossy compression schemes as, for example, JB2 from DjVu.
We can see in the picture below a lot of noise and flipped pixels at the
left side, where the black-and-white text has been produced by application
of the Floyd-Steinberg error diffusion method, while at the right side
it was produced by the nearest colour reduction method. The text is taken
from a sample page A4 scanned at 300 dpi. Of course, other error diffusion
methods can give other results: in this case, the Stucki error diffusion
method produced less noise (see attachments).
It is also interesting to observe that the size of the uncompressed 1-bit sample
file from the beginning of this study will have only 100 KB - comp. with
2.4 MB of the true colour image. This fact is worth to be considered seriously
As we can see, sometimes, it is reasonable to refuse colours and to digitize
in shades of grey: usually it means to create a palette of 256 shades of
grey that is a direct consequence of conversion of the true-colour (24-bit)
image into shades of grey (8-bit). It is interesting to realize that even
through such a conversion we reduce the size of the image file three times.
We may also reduce the number of shades of grey in some cases onto 16.
This does not seem to be so disturbing as in case of a palette consisting
of only sixteen colours.
This may be perhaps explained by the fact that we have been used to
read bad quality newspapers. However, everyone must decide on his or her
In other words, we may reduce the true-colour image into the 256-palette
image with application of two different operations:
The difference is that almost each 256-colour image has another colour
palette so that it is almost impossible to combine such images. On the
other hand, the grey palette is the same for all 8-bit images; therefore,
such images are better comparable with one another. Evidently, in many
cases it is preferable to scan in 256 shades of grey than in 256 colours.
Of course, it may be also said that the grey palette image is a special
case of a colour image. This can be easier explained by application of
RGB or HSL colour models, but this is not the purpose of this article.
conversion into the colour palette
conversion into a palette composed from shades of grey
Compression can be lossless or lossy; data delivery in library services
is mostly interested in lossy compression schemes. Traditionally, the development
of compression schemes happened in two ways:
It has been demonstrated in former studies how efficient lossy compression
schemes in these two areas can be today. To be correct, it must be said
that some progress has also been achieved in lossless compression that
is more evident in 1-bit compression schemes.
compression of 1-bit images
compression of colour images (all that is not 1-bit image is colour image)
Best lossless and
lossy standards and their emerging follow-on
The best-standardized lossless compression schemes used today are CCITT
Fax Group 4 for 1-bit (usually implemented in TIFF) and PNG (usually implemented
in PNG format). It is also known that the 1-bit ISO compression scheme called
JBIG is better than CCITT Fax Group 4, but unfortunately, it is not used.
Final ISO draft for JBIG2 is on the way to be approved, but the question
is, whether these schemes will be used in products at affordable price.
Simultaneously with JBIG2 initiatives for 1-bit images, JPEG2000 initiatives
for colour images are under development.
The performance of JBIG2-related encoders is evidently far better even
in lossless domain than the performance of CCITT Fax Group 4. The best
1-bit encoders implemented in deliverable products are JB2 developed by
AT&T and LDF 1-bit encoder developed by the German Luratech company.
Their performance is even more accentuated when enabled also for lossy
compression as also demonstrated in our recent study.
In the domain of colour image, it seems that lossless wavelet compression
techniques will not outperform PNG with more than 10%; therefore, we can
frequently hear that PNG is a format that has its future well ahead. However,
in the lossy domain, wavelets will replace the best-standardized lossy
compression scheme of today that is DCT (discrete cosine transform) as
known from JPEG.
||today best performance
||CCITT Fax Group 4/TIFF
It is also true that PNG and JPEG DCT compression schemes can be applied
in TIFF, but these options are not used. There is almost no accessible
software to encode TIFF in such a way - Graphic Workshop Professional by
Alchemy Mindworks can do it - as well as there is almost no viewer to work
with it. TIFF also admits the LZW compression scheme that we know from
GIF, but the efficiency of LZW is inferior face to the PNG scheme.
The key problem of any lossy compression is establishment of the quality
factor or threshold limit of acceptability of the compressed image for
the given purpose. This is always individual and the purpose plays here
the very decisive role.
The following table sums up the options available in colour reduction
and compression domains. We mention only the most used standards and approaches.
||uncompressed size in bytes
||applicable lossless compression
||applicable lossy compression
|24 bit (true colour)
DjVu/IW44, LDF/LWF, MrSID
PNG and TIFF/LZW also possible
|PNG, GIF, TIFF/LZW
PDF, PNG, GIF, …
LDF/LDF 1-bit encoder
As we can see, the reduction of the colour depth is a mighty tool for
the document delivery services.
The 8-bit image is usually dominated by the GIF format that uses the
LZW compression scheme. However, we can encode such an image also in other
formats as, for example, TIFF - using the same LZW compression scheme or
the Packbits scheme - or PNG. The smallest files are obtained in PNG.
The same can be said about 4-bit image. For 4-bit and 8-bit images lossy
compression schemes are not used.
The 1-bit image, however, is a special case - see our former study about
efficiency of 1-bit compression schemes.
The purpose of library services is efficient delivery of highly relevant
information with the volume of data as reduced as possible. Readers are
interested in scanned pages of documents and these documents usually consist of
text and images. It is said that their content is mixed.
The problem is that the most efficient compression of text is possible
in 1-bit compression schemes, while the rest of information must be delivered
in colour or shades of grey. It is evident that if we achieve to compress
separately the text and separately the colour from an image, we can probably
have better lossy results than with any other techniques. Even if we combine
CCITT Fax Group 4 and JPEG (as it is possible, for example, in the MRC
format designed by Imagepower Inc.), the result is more efficient than
that achieved with JPEG only.
It is evident from former tests that the best combination would be a
lossy 1-bit compression scheme combined with a good wavelet compression
How to segment the
However, how to automate the segmentation of objects into those to be
encoded bitonally and those to be encoded in colour. Theoretically, there
are two possible approaches:
vertical (in-depth) segmentation
take a page from a modern journal, we can observe that there is text and
that there are images and simple graphics; moreover, there is also coloured
background or images used as background. The idea of horizontal segmentation
was demonstrated by Power Compressor by Imagepower Inc.
The segmentation - called here zoning - could be performed automatically
and possibly corrected manually or it can be done only manually. The applicable
segments are text, simple graphics, and image (photograph). When segmented,
the page consists of areas marked as T, G, or P; these areas can be also
The areas are then encoded separately and merged into one format
called here MRC (standing for mixed raster graphics). The user can set
up, which compression schemes will be applied for each area type. For 1-bit
schemes, the choice can be CCITT Fax Group 4 or JBIG2, for graphics and
image JPEG or a wavelet scheme, which was here a proper WI format - now it is JPC format that is the Imagepower vesion of JPEG2000. The user
can also set up the quality parameters of some of these schemes. The text
background can be also set up to zero.
When testing the format, we encountered a lot of problems due to the
complexity of applied areas in more diversified pages. The format worked
in case of simpler pages only. The browser plug-in was unable to display
the results correctly, while the plug-in for the higher version of the Power Compressor products was still under development.
Other horizontally oriented solutions have not been found.
The vertical segmentation tries to split the image into layers situated
one above another. The layer in the foreground is compressed bitonally
and the remaining layers in colour. Due to the fact that the segmentation
is performed automatically, some pixel groups - be they text or not - are
taken as foreground to which colour can be added in the second compressed
layer. The rest is considered image and it is compressed in true colour.
Keeping in view that the character of individual images differs and
that not every image looks well when compressed as mixed raster content,
authoring tools of today formats have two other options: compress as text
or compress as photograph. In these two cases the wavelet or the bitonal
compression encoders are disabled.
We can see here the difference between the correct use of the mixed
raster content format for encoding a photograph and the incorrect use.
The left image was encoded correctly, because it was not segmented into
layers and it was encoded entirely as a background image (only true colour
On the other side, the right image was not encoded by a method that
was appropriate for it. It was compressed as text and image, i.e. segmented
into 1-bit and true colour areas.
The left image is one image in one, but the right image can be split
into foreground, coloured foreground, and background image.
The segmentation looks like this:
In this explanation, the LuraDocument format was used to demonstrate
the character of the mixed raster content encoding.
Another used format for this type of encoding is DjVu, it works on the
same principles, but the encoding schemes differ.
When the LuraDocument uses for 1-bit encoding its own 1-bit encoding
scheme (TIFF G4 option is also possible), DjVu uses its own JB2 encoding
scheme that has a different performance. When for image in LuraDocument
the LWF (LuraWave) format is used, in DjVu it is the IW44 format.
The situation of these two formats is quite interesting, because our
former tests have shown that as to 1-bit encoders the DjVu JB2 is better
than the LDF 1-bit encoder, but LWF on the other hand is better than IW44.
However the word 'better' in case of 1-bit encoders means smaller JB2 files
of high quality, while the same word in case of the true image means the
better LWF image quality at the same size. It is evident from here that
if LDF wants to match DjVu, it must reduce the image background size and
quality. This may be a problem, because the quality difference between
IW44 and LWF does not play so important role in mixed raster content delivery
as it does play in true image representation (representation of photorealistic
The user expects reduced quality in case if mixed raster content or
he can admit it. However, in the case of true image, he requires and expects
to receive better quality, because the role of the image in mixed raster
content is to illustrate the text, while in the case of true image, the
image represents itself.
When IW44 scheme segments the image into four sub-layers, LWF apparently
uses other techniques, because it is faster at encoding as well as at decoding
(display). At larger images of more than ca. 50 MB uncompressed image,
the freely available encoding tools LuraDocument Capture and DjVu Shop
work differently. When the DjVu Shop is unable to encode such large images on less powerful computers,
the LuraDocument Capture has no problems and it is relatively fast. This was the situation with 64 MB RAM, but when upgraded onto 128 MB or more, the DjVu Shop proved a good reliability and acceptable speed.
The LuraDocument Command Line Tool is even faster and our tests have shown that the DjVu Command Line Encoder on a Linux machine is also very fast.
However, which of these tools offers the best compression? For the moment
being, we know that for 1-bit images it is DjVu and for true colour images
LDF. These results remain outside of the mixed raster content encoding
area, but they can help us when we start to decide, which modern tools
to use for our data delivery. In this sense, it can be also interesting that the Luratech Company has recently implemented JPEG2000 into LWF, version 3.0, into their special JPEG2000 programme. Together with the Image Power Inc. they seem to be the first companies, which have started to develop deliverable tools with JPEG2000.
Testing a mixed raster
content page delivery
It is evident that the efficiency of the mixed raster content compression
and delivery will depend on the segmentation of the image. To discover
this, we scanned a page from a journal with text and colour images at 300
dpi that resulted into a 24,053 KB uncompressed image.
We used the default values of the encoders, which were DjVu Shop, version
2, for the DjVu format and LuraDocument Capture, version 1.1, both free
tools for non-commercial use.
The default values for the DjVu tool were 300 dpi for text and 100 dpi
for image, while the text background was compressed with the low resolution
of 25 dpi; the image background quality factor was set up to 75 and the
text was compressed by the lossy scheme. The DjVu output file had the size
of 68.1 KB.
When the text was compressed as lossless mask, the resulted DjVu file
was 81.4 KB. We also tried to apply the aggressive mask for text, but there
was no further gain in the final file size compared to the lossy scheme
- perhaps only several bytes that was not so important. When we set up
the text background mask to 100 dpi, then the final DjVu file size was
We also combined the set-up in the manner as follows: 300-dpi mask for
text and lossless compression, 100 dpi for text background, and 100 dpi
for image background. In this case the final size of the DjVu file was
The default values for the LDF tool were 'catalogue medium' as to image
quality and 300 dpi, while the bitonal encoder was LDF 1-bit encoder. The
obtained LDF file had 160.5 KB. When we set up the LDF tool to 'catalogue
low' - this referred to the background image - then the LDF output file
was 123 KB.
We also used the LuraDocument Command Line Tool whose testing was enabled
for us by the Luratech Comp. and we tried to combine encoding masks of
various resolution values separately for text and image. When we used 300
dpi for text and 100 dpi for background image, the final size of the LDF
file was 99 KB, but the quality of the encoded image was rather poor.
Another possibility was to omit the text background layer and to encode
the text bitonally and the image in LWF. The final size of the LDF file
was then 80 KB, but there was loss of that information, which we did preserve
in DjVu; therefore this result could not be used for comparison.
Finally we took the CCITT Fax Group 4 scheme for text in place of the
LDF encoder. The final size of the LDF image grew with 34 KB, i.e. 152.87
KB (G4 and medium LWF quality), while for LDF 1-bit and low LWF quality it was 121.84 KB.
We tried also other settings for the resolution of both masks, but we were
unable to create smaller files with acceptable readability or understandability
especially for graphics.
Comparing LDF with DjVu
We can sum up the results into the following table:
||300 dpi lossy
(cannot be set up)
|300 dpi medium
||NA (cannot be set up)
As we can see, the LDF files are larger, the lowest difference being
ca. 30 KB, but in this case the LDF file had a very bad image quality and
it was not comparable to the 68.1 KB DjVu file. On the other hand, the
default output images of 68.1 KB for DjVu and 123 KB for LDF were well
comparable; it can be said that they were equal as to their quality when
evaluated from the user's point of view. Of course, they were not identical
as to various textual and image artifacts and the quality of representation
of certain elements of the scanned page, but they could play the same quality
role in the delivery services.
Analyzing the layers
As demonstrated on the comparison of both DjVu and LDF segments (see
attached Comparison of DjVu and LDF layers), it is evident that
here the LDF 1-bit segment carries more information than the JB2 segment.
This situation does not change if we add the text background to have the
whole coloured foreground. On the other hand, the LDF image (LWF format)
background is more or less another text background, while the DjVu image
(IW44 format) can be considered as a certain kind of blurred image.
In general, the LDF and DjVu formats are quite comparable as to size
in the background image representation, where the quality aspect would
be slightly in favour of LDF (LWF).
However, LDF stores here fewer data in the background image segment
than in the foreground image segment, while DjVu does it on the contrary.
If LDF managed to store the same or comparable segments of data in the
same layers as DjVu, DjVu would outperform LDF only in the 1-bit domain
thanks to the demonstrated quality of JB2 and the difference in size of
LDF and DjVu files would not be so big.
Unfortunately, LDF stores substantially more information in 1-bit layer
than DjVu; therefore, the sizes of the output files are so different: DjVu
outperforms (in the case of our page) LDF twice. Apparently also the resolution
of the background image is higher (300 dpi) at LDF than at DjVu (100 dpi).
On the other hand, if we reduce the resolution of the background image
at LDF down to 100 dpi (Command Line Tool), we are almost losing the image.
Maybe if LDF would try to store more information in the background image
than it does now, it would not need to store so much data in the 1-bit
layer, where its performance is not so good. Perhaps thus the difference
between DjVu and LDF would not be so acute as it is now. To achieve this
more balanced situation, the vertical segmentation of LDF should be tuned
more in favour of the LWF compression scheme (background) than in favour
of the 1-bit scheme as it is now.
Today, LDF stores less information than DjVu in the segment in which
their performances are comparable and it does store more information in
the foreground segment, where it is not so efficient, because DjVu/JB2
outperforms the LDF 1-bit encoder substantially (see our former tests).
||DjVu/JB2 is substantially better
||LDF/LWF is very slightly better
with traditional approaches
If we take the problem from a very practical point of view, we can compare
the efficiency of reduction of the colour depth and application of various
lossless and lossy compression schemes. All figures in graphs are given
True colour - 24-bit image-
16,7 million colours
Our test image of more than 24 MB was scanned from a colour journal
at 300 dpi and compressed as true colour image with various lossless and
lossy compression schemes with the following results:
We have applied the wavelet compression only within DjVu and LDF as a component
part of the MRC approach.
the best lossless compression scheme for the true colour image was PNG
in the PNG format; in the Graphic Workshop we were even able to set up
a tougher compression indicated as PNG max., the difference between this
format and TIFF/LZW is substantial. In plus, PNG is recommended for the
World Wide Web, while TIFF is not;
in the lossy domain, JPEG proved its reliability and the savings were substantial;
the MRC-based formats DjVu and LDF were able to compress the images even
more while preserving good colour representation and readability: 123 KB
in LDF and only 68 in DjVu are incredible. Very recently, we have also tried
the new DjVu 3 format in the DjVuSolo compressor - next version of the
DjVuShop 2 - and we have been able to go down to 61 KB only.
of colour depth
The colour depth of the test image was reduced onto 256 colours and
after it onto 16 colours. These two images were compressed in GIF and PNG.
In both 256 and 16-colour domains, the PNG format was more efficient than
GIF, but the files remained rather large: 4790 and 1940 KB. It is evident
that the lossy true colour JPEG is a better solution than reduction of
the colour depth.
We reduced the number of colours onto black and white only and compressed
this image with several lossless and lossy methods. The importance of TIFF/G4
was confirmed as well as very good results of LDF and DjVu/JB2. In the
lossy domain, DjVu/JB2 was able to compress the file down onto 61 KB, while
the DjVu 3 onto 56 KB only. The 1-bit source image was obtained through
dithering with the nearest colour reduction method so that it may be said
that all the obtained images are very suitable for OCR processing.
It is evident that for document delivery the best solution is the true
colour mixed raster content approach. Thanks to this, the user has access
to good quality colour information and well readable text. The difference
between the best 1-bit compression performance and the best mixed raster
content performance is very small. For LDF it is only 6 KB (117 KB vs.
123 KB) and for DjVu only 7 KB (61 KB vs. 68 KB).
However, within the most used DjVu MRC settings such lossy 1-bit compression
- that is default - and lossless 1-bit compression, the size of the MRC
image is smaller than that of the corresponding 1-bit image. This fact
is rather interesting and it is due to optimized segmentation of layers
and their distribution to 1-bit and colour compression areas. As it can
be seen from the attached samples, the wavelet compression takes a part
of data from the traditional 1-bit area compared with the pure black-and-white
This was also possible thanks to the fact that the compressed original
page was composed from about 50% of text and 50% of images.
in shades of grey
However, in our digitization programme, acid-paper materials must be
frequently scanned in grey scale, because the condition of the original
is so poor that through dithering onto 1-bit image the readability of the
text can be seriously endangered.
To test the compression also here, we have taken an image of a typewritten
page from the so-called Russian Historical Archives Abroad. The JPEG image
in 256 shades of grey has 994 KB - face to the uncompressed image of ca.
22 MB - while the JPEG with the Quality Factor 90 has 517 KB and that with
the Factor 95 has 350 KB. The text has remained well readable, but the
background is full of artefacts caused by the tough compression.
The LuraDocument file - LuraDocument Capture software with the low quality
background - has produced a file sized 321 KB; the result was excellent.
We have also tried to dither the image onto 1-bit and to compress it afterwards.
We have thus obtained a TIFF image compressed with CCITT Fax Group 4; it
has the size of 184 KB. We have been able to compress it down with LDF
onto 134 KB, with DjVu 2 onto 90 KB, and with DjVu 3 onto 87 KB. We have
also tried to compress and to dither the image with DjVu directly from
JPEG: the resulted file is slightly larger in this case - 117 KB.
The third part of our tests has been application of the DjVu MRC compression
scheme. Having in mind the MRC result obtained in LDF - 321 KB - we have
not been surprised with 296 KB obtained through DjVu 2. However, the surprise
has come with DjVu 3 (DjVuSolo software), because the file has been compressed
onto only 153 KB that is a fascinating result.
Also in this case, we have observed that a part of the text layer has
moved in the MRC scheme into the wavelet compression background segments.
In DjVu 3, the 1-bit file has 87 KB, but the 1-bit layer in the MRC solution
has only 80 KB.
It is evident also that the methods for simple dithering try to preserve
the maximum of information in the black-and-white representation. The mixed
raster content approach places, however, only a part of it into the 1-bit
text layer, while all the rest is encoded in wavelets and very efficiently.
The new version 3 of the DjVu format as tested in the DjVuSolo program
enables a very efficient compression, but it is much more computation intensive
than the former version. The DjVuSolo program also requires display of
the source image and only after this it is possible to convert it, while
the DjVuShop loaded the source image through conversion.
The new program offers also less opportunities to set up the compression
ratios than the old one does. It has been designed for practical use and
as such it seems to copy the philosophy of the LuraDocument Capture. However,
the LuraDocument Capture program is very fast, while the Windows DjVu encoding
tools are slow on less powerful computers. They require more memory and better processor performance.
However, the professional command line tools on the Linux platform
are substantially faster and also the variety of possible settings is very rich. The results of the DjVu encoding are astonishing;
therefore, we have decided to apply them in our digital library.
There are a lot of opportunities for delivery of scanned materials over
networks. The word 'delivery' supposes existence of a contract or some
basic understanding between the service provider and its user. If this
contract enables reduction of data down to the limit of acceptable readability
and understanding, then we can apply the mixed raster content (MRC) delivery
concept. However, we must be aware of the fact that MRC formats have a
lot of parameters and that we must know their properties to be able to
apply them correctly.
There is always difference between a page of a journal with text and
images, a page from a manuscript or a photograph, or a simple black text
on white paper. All of them can be delivered by DjVu or LDF, but not onto
all of them the MRC philosophy can be applied.
The MRC formats are an opportunity of application of two different compression
schemes onto the same image. They give us also the liberty to switch off
one of the compression schemes and to use that, which remains. In these
cases - when one of the compression schemes is switched off - they behave
as very efficient - the most efficient ones of today - formats for 1-bit
or colour encoding. They outperform traditional 1-bit and photorealistic
compression solutions. Furthermore, they are even much better when we allow
them to apply both compression schemes in one image.
It is evident from the tests that if we want to economize storage space
and to reduce the network traffic, DjVu is a better solution than LDF.
On the other hand, it should be said that both solutions are on the market
and that they are better than any third solution that may be applied.
If we work with free tools, we have comparable viewing plug-ins for
the two most used web browsers and we have two Windows tools to encode
When the files get larger, we will appreciate the speed of the LDF tool on less powerful machines.
When moving towards encoding of larger files, the DjVu Shop requires more memory. However, when viewing the files in web browsers, the DjVu plug-in seems to be faster and to offer more comfort. Its speed is also partly given by the characteristics of the DjVu image for the first display of which only a part of data is needed.
Both formats enable storage of more images (scanned pages) in one file.
The free LDF tool enables creation of up to eight images in one file and
the license to open the limits of this functions costs only ca. 30 USD.
If we want to do the same with DjVu, we can use the DjVuSolo program and, of course, commercial tools or special free command line tools for creation of multi-image files that may be very difficult for common users.
On the other hand, the DjVu philosophy and the free command line tool for
grouping individual DjVu files enable virtual - not physical - merging
of separate files into one. The advantage of this solution is that only
the page requested during browsing by the user is transferred over networks;
not all the files or pages as it happens in case of LDF, PDF, or multipage
On the left image all the individual files are transferred in one file.
This method is used in PDF, LDF, and TIFF. We have this option in DjVu,
However, in case of DjVu, the total size of individual 1-bit layers grouped in
one file is smaller than the total size of these 1-bit layers remained
in individual files ungrouped. This is due to use of the common shared dictionary
for all the pixel groups gathered in the former case, while in the latter
case several individual dictionaries are used.
On the right image, we receive - when clicking on the name of the metafile
- only this very slim metafile and the first page. When we follow to browse
the document, we receive only that file, which we require: it may be the
previous file or the next file, but also any other file referenced. We
only indicate in the plug-in menu, which number we want.
If we want to use LDF or DjVu in high volume, even in some document
delivery application in the digital library, we will need the so-called
command line tools, which are very fast. The LDF command line tool costs almost 3000 USD, while
the DjVu command line tool for mixed raster content costs 7000 USD. This
fact can be also important when taking our decision, which of them we should
Another important factor is that, which input formats for conversion
are applicable. The Windows tools accept JPEG, while the command line tools
- in case of LDF - not, while the DjVu command line encoder does accept
it. LDF command line tool accepts uncompressed bitmaps as TIFF, BMP, or
RAS. It is evident that JPEG input is important for some segment of use.
This is true especially when we convert existing files and do not scan
only for delivery, because in this case it is rather probable that storage
of true colour image is done rather in JPEG than in some uncompressed format.
The following samples are available via Internet:
is deputy director of the National Library of the Czech Republic, responsible for strategic planning, research, and technology development. Since 1993, he has been active in development of co-operation with the UNESCO Memory of the World programme; today he guarantees the national programmes Memoriae Mundi Series Bohemica as a programme of digital access to rare documents and the Digital Library programme as an archive for the digital documents produced by Czech libraries. In 1990s, he was active also in various other initiatives launched by the European Commission and UNESCO. In 1997-2000, he was member of the International Advisory Committee of the Memory of the World programme and since 1996 he has been member of its Sub-Committee on Technology.