I Miss my Viewfinder !

On 03/06/2011 10:28, Floyd L. Davidson wrote:

Here's the entire sentence, and now *you* can tell us
just what difference the rest of the sentence makes!

"There really isn't any difference, except that
signal is compressed according to an algorithm that
can reversibly restore the original uncompressed
signal, whereas applying that same algorithm to truly
random data typically just produces more useless
data."

He is slightly wrong. Applying the same algorithm to truly random data
fairly quickly generates an illegal decode symbol error in most codecs.

That is essentially what happens on digital TV or DAB when noise in the
stream crashes the decoder and it breaks up into Picasso block art and
tries to burn out the tweeters with chirped ultrasonic clicks.

Noise is not data and data is not noise.

Although that is true. It is also true that an important characteristic
of truly good random stream of bytes is that it should incompressible by
any lossless algorithm (and it may even generate a larger file).

What he says is also nearly true too for a JPEG stream. The encoding
algorithm of JPEG specifies FF as and escape and FF,00 as a true FF in
the stream and so is slightly imperfect even when Huffman optimised.

There is usually a measurable bias towards symbols with 3 or more
consequtive '1's in the byte with default encoding tables.

Although the results are strongly data dependent you have to be encoding
an image of black cat in a coal cellar or igloo in a snowstorm before
the encoded data looks noticeably different to random bytes.

A typical JPEG image off my camera is ~4.7MB and default Huffman tables
has a bytewise entropy of 5.49 - it could be improved by optimising the
tables but few cameras do. The theoretical maximum is ln(256) = 5.545

This idea that if it looks the same on an oscilloscope
then it is the same is hilarious; but wrong.

Unless you know the algorithm to decode it then any well encrypted or
compressed data looks very much like random noise. If it didn't then you
would be able to find another algorithm to make the file still smaller
until you had something that was truly incompressible.

Bytewise entropy isn't foolproof. It would give the same entropy to both
of the following because it doesn't look at conditional symbol
probabilities. One is clearly a lot more compressible than the other.

AAAABBBBCCCCDDDDEEEE

ABCDEBDCEAADBCAEEBCD

The main utility of bytewise entropy or compressibility is that it is a
good heuristic test for the validity of a JPEG stream. Basically if it
will compress by more than about 5% and it is supposed to be a real JPEG
photographic image (as opposed to a pathological testcase) then the file
is seriously damaged.

Regards,
Martin Brown

Mxsmanic wrote:
Martin Brown writes:

It will be incompressible if it is infinite in length. A finite length of
random data will be compressible by some algorithm, based on the statistics of
the stream.

Sure, put the random data into the algorithm.

probabilities. One is clearly a lot more compressible than the other.

AAAABBBBCCCCDDDDEEEE

ABCDEBDCEAADBCAEEBCD

They should both be equally compressible, but not by the same algorithm.

Provide a general algorithm then.

Basically if it
will compress by more than about 5% and it is supposed to be a real JPEG
photographic image (as opposed to a pathological testcase) then the file
is seriously damaged.

That's why I always shake my head when people send me a JPEG in a ZIP file.
Ditto for MP3 or most types of video.

Error detection.

-Wolfgang