Documente Academic
Documente Profesional
Documente Cultură
JPEG2000
millennium
jpeg
Su r fing the Web will never be the same again. Wavelet compression is set to ch a n g e
the way we view, transmit and edit images. Used in the new JPEG2000 st a n d a rd ,
this technology shrinks image data to a fraction of its original size – and enables
f a st decompression without loss. Here are the facts .
Figure 2: A Web-based wavelet-compressed image appears quickly, then gradually gains clarity.
(think of the squiggly waves on an oscillo- applying the right wavelet transformation In simple terms, this is how such an image
scope) could be expressed as the sum of a to a photographic image can yield excellent would be decompressed:
series of sine and cosine functions with compression without the blocky artifacts
particular parameters. The procedure that found in highly compressed JPEG images.
renders a periodic function into such a The biggest diff e rence between the 1. Display the stored minimum size
series is called a Fourier transform, and is wavelet transformation and the direct cosine image, but stretched to the final full
useful for enabling scientists to see pat- t r a n s f o rm used in J P E G c o m p re s s i o n width and height. The image will look
terns in the transformed data that weren’t involves scale. The DCT is always applied quite fuzzy.
evident in the raw data. to 8 by 8 squares. In contrast, the wavelet 2. Repeatedly apply the sets of
The Fourier transform applies to mathe- t r a n s f o rmation is applied to the whole differences, doubling the image’s
matical functions, not to digital images, but image repeatedly at different scales. The clarity at each step.
a related transformation called a discrete wavelet functions alternately record and
cosine transform (DCT) can be used on remove detail from the image. Most wavelet Figure 2 shows an enlarged view of four
non-continuous data. A DCT transforms an functions are too complex to describe eas- stages in the process of decompressing a
array of data into a set of parameters that ily, but the Haar wavelet, one of the first simple, 32x32-pixel gray scale image of a
can be used to re-create the original data invented, can be approximated in a simple five-pointed star. The image at left is
precisely. JPEG compression breaks an description: stretched from just 4 by 4 pixels and does-
image into 8x8-pixel blocks and applies n’t show much detail. When the first set of
the DCT to the pixel colours in each block. differences is applied, the next image is the
There’s no loss of data at this point, but 1. Reduce the image to half its width result; this is effectively an 8x8-pixel image.
there’s no compression, either. and height by averaging the colours of Applying the next set of differences yields
The JPEG algorithm reduces the space each 2 by 2 block of pixels into a the third image, which is much clearer. Final-
required to store the image by quantising single pixel. ly, the last set of differences restores the
the output of the DCT – dividing the para- 2. Record the differences between the original 32x32-pixel image, the rightmost
meters by specific values and rounding the original pixels and the averaged pixel in the figure.
result to the nearest integer. This step that replaced them (this will often be In the real world of wavelet compression,
reduces many of the parameters to zero. zero or close to zero). functions much more complex than the
Also, there are fewer distinct values, which 3. Repeat until the image is reduced Haar wavelet are used. Still, this approxi-
makes the data stream more easily com- to a specified minimum size. mation should illustrate the principles
pressible. Some data is lost, but the quan- 4. Store the minimum size image. behind wavelet compression.
tising step is designed to minimise the visu- 5. Store the sets of difference infor- Wavelet compression is extremely well
al effect of that loss. The resulting data mation in reverse order. suited to Web-based images. The moment
stream is compressed again, using a lossless the minimum size image has been received,
algorithm that assigns the smallest codes to At each stage in this process, the image a fuzzy picture can be displayed. As the rest
the most common bit patterns. detail is removed by averaging adjacent of the data stream arrives, the picture gets
pixels, and recorded by storing the differ- steadily clearer. Also, there’s no need to
WAVELET COMPRESSION ences. The resulting data stream includes maintain copies of the image at different
Wavelet analysis has been around for almost many zeroes, representing pixels that are sizes. To obtain a half-size image, you sim-
100 years. As with Fourier analysis, scien- the same colour as their neighbors. Now, ply begin the display process at half the
tists use wavelet analysis to find patterns in as with original JPEG, a lossless compres- original size and omit the final set of differ-
data. The technique has been used to clean sion algorithm that works well with such a ences. No doubt about it, wavelet com-
up “noisy” statistical and audio data, includ- data stream is applied. Additional com- pression is going to improve your browsing
ing an antiquated recording of composer pression can be gained, at the expense of experience; you’ll see better pictures on the
Johannes Brahms playing part of his own faithfulness to the original, by taking val- Web, and they’ll appear faster than ever.
Hungarian Dance No. 1. Wavelet analysis ues near zero and changing them to zero. Visit www.lizardtech.com or www.lurate-
has applications in mathematics, quantum Visually important elements like edges ch.com for currently available examples, or
physics, electrical engineering and seis- won’t be affected, because there the dif- wait a little while longer for products imple-
mology, among other fields. As it happens, ferences are large. menting JPEG2000.