compression of the image adolf knoll national library of the czech republic
TRANSCRIPT
Compression of the image
Adolf Knoll
National Library of the Czech Republic
General schemes for application of compression
The schemes adapt to the character of the represented objects:
Bitonal image (1-bit, black-and-white) Colour photorealistic image Mixed document (two above-mentioned
components)
Trends
Bitonal from CCITT Gr. Fax 3 and 4 to JBIG variants
Photorealistic Lossless compression: PNG, TIFF/LZW Lossy: from JPEG DCT to wavelet
Mixed document Both applied (Mixed Raster Content –
usually vertically)
How is it built into formats?
Trying to have it in ISO TIFF (even JPEG, LZW, or PNG) – but it is not enough due to lack of tools for conversion and display.
That is why the other more suitable formats are used: JPEG, PNG
That is why there is a lot of development in the area of mixed formats – they do not aim to become ISO
Relevant directions
Bitonal image JBIG2 (ISO) – no support (exc. Xerox), but
many similar activities Photorealistic image
wavelet JPEG2000 and many other non-ISO initiatives (WI, LWF, IW44, SID, Imagepower IW, …)
Mixed content DjVu, LDF, Imagepower MRC
Aims
Image Archiving standardized
archival format (TIFF, JPEG, PNG, …)
Image Delivery More efficient
modern format (JB2, MrSID, DjVu, LDF, …)
Which relationship will be between both of them?It will be defined by the goal of the project.
Around compression
Pre-processing of the image Compression Encoding in a format De-coding from the format De-compression Display – print-out
Pre-processing of the bitonal image - I
Efficient schemes are built on possibilities to apply vocabularies of pixel chunks/groups: E.g. a text is an image that can be interpreted as
several dozens of images of letters, while the repeated occurrence of each letter can be represented by its coordinates (x,y) and reference to a dictionary in which there is only one representation of similar letters (digitized only once as a bitmap)
This method is called PATTERN MATCHING, but…
Pre-processing of the bitonal image - II
However, scanned texts have a lot of information noise in individual pixel chunks representing, for instance, letters in text
Therefore, it is convenient to reduce differences between identically indentifiable chunks smoothing pixel flipping noise removal
Smoothing and pixel flipping
Problems in pattern matching
Česká republika
Low quality original and/or scan + inappropriate processing
Soft pattern matching
Better work with dictionaries; replacement only there, where the threshold value of the pixel chunk is satisfied
If not, the whole small bitmap is stored Tuning of these mechanisms is a key
to successful application of the lossy compression of a bitonal image.
How to know…
Libraries have documents of various qualities- also very bad
These documents are more difficult to process than good samples presented by software producers
Tests… tests… tests… on typical materials
Bitonal compression
Lossless (LZW, PNG, …, CCITT Fax Group 3 a 4, JB2, JBIG, JBIG2, Algo Vision/Luratech (1-bit LDF component)
Lossy modern schemes: AT&T (Lizardtech) (JB2) – soft pattern
matching ImagePower Inc. JBIG2 (JB2) – only pattern
matching Summus Inc. (Lightning Strike), ...
GIF would beslightly worsethan PNG
Květy české – 19th century Czech journal
Impact of the quality of digitized originals on performance of compression schemes
JB2
Most efficient compression schemes JB2 from the DjVu format (AT&T).
It enables compression: lossless lossy aggressive – while preserving high
quality
JB2 as a component part of the DjVu format
More files can be merged and saved into one (as PDF) – they have the common dictionary so that together their size will be smaller than the sum of all individual files
More files can be virtually joined (they are called one after another from the server)
More advantages: display, references, OCR, … (DjVu plug-in)
Expensive or free software for Linux or Solaris
Samples and résumé
Monitor and test new approaches for image processing
They can be very suitable for document delivery services Image servers Scanned content CLICK!!!
Which formats to use for bitonal image?
If you have no special tools: GIF
If you wish smaller files, use PNG Both are recommended for WWW However, TIFF/CCITT Fax Gr. 4 is
better Use DjVu, if you wish very small files
Problems
Good image editing software does not support TIFF with Gr. 4 encoding
Display possible within normal Windows tools
GIF and PNG support also higher brightness resolution (8-bit / 24-bit) – take care not to save bi-level image in higher image depth
DjVu – necessary to solve authoring software problem
Lossy compression – bitonal image
Compression of colour images
Lossless LZW
GIF (8-bit only) TIFF (5.0)
PNG Wavelet
JPEG2000 (JP2) …
Lossy DCT (JPEG) Fractals Wavelet
IW44 LWF, WI JPEG2000 (JP2) MrSID, …
Classical (LZW, RLE, DCT) versus wavelet approaches.
True colour image
DCT
wavelet
Testing compression efficiency
Sample Reference Full-colour (JPEG, wavelet) 1-bit (establish tresholds – Paint Shop
Pro, LuraWave) MRC (same sample – DjVu Solo)
Compression efficiency – bitonal image
Compression efficiency
True colour
How to apply compression?
It depends on the character of objects in the image: Photorealistic image (JPEG, wavelet) Text and simple blac-and-white graphics (Fax
Group 4, JB2, …) Colour graphics (problem to compress with losses
– better lossless PNG or GIF – application area of vector graphics - SVG)
Mixed content (composed solutions: DjVu, LDF, …)
The most efficient solution
To segment images into two or more groups of objects:
1. Objects good for bitonal conversion
2. Objects good for true colour representation
Tto compress each group separately and then merge into one format.
Horizontal segmentation/zoning
Horizontally- Text- Grafics- Photographs
Imagepower Inc.
Vertical segmentation/zoning
Vertically Foreground Background
Lizardtech Inc. (AT&T)Luratech GmBH
DjVu, LDF
Comparison of DjVu and LDF
DjVu
6 layers
Foreground: JB2 IW44
Background: 4 layers IW44
LDF
3 layers
Foreground: LDF 1-bit Comp. LFW
Background: 1 layer LWF, JP2
Bitonal versus composed image
Grey level
Other DjVu properties
More images in one: as TIFF, PDF, LDF, …, with use of the
common dictionary of pixel chunks Virtually: pages remaion on server and
only that page that is called is delivered
Multiresolution image
MrSID In one file several (up to 8) images in
various resolutions Sample Efficient with an image server
SAMPLES
Samples of various compression solutions