39078

神经网络昆斯林的新闻组分类2006

What it does get rid of is deliberate information loss in the coefficient
quantization step.  There is still a good deal of information loss in the
color subsampling step.  (With the V4 free JPEG code, you can also say
"-sample 1x1" to turn off subsampling.  Keep in mind that many commercial
JPEG implementations cannot cope with the resulting file.)

Even with both quantization and subsampling turned off, the regular JPEG
algorithm is not lossless, because it is subject to roundoff errors in
various calculations.  The maximum error is a few counts in any one pixel
value; it's highly unlikely that this could be perceived by the human eye,
but it might be a concern if you are doing machine processing of an image.

At this minimum-loss setting, regular JPEG produces files that are perhaps
half the size of an uncompressed 24-bit-per-pixel image.  True lossless JPEG
provides roughly the same amount of compression, but it guarantees
bit-for-bit accuracy.

If you have an application requiring lossless storage of images with less
than 6 bits per pixel (per color component), you may want to look into the
JBIG bilevel image compression standard.  This performs better than JPEG
lossless on such images.  JPEG lossless is superior to JBIG on images with
6 or more bits per pixel; furthermore, JPEG is public domain (at least with a
Huffman back end), while the JBIG techniques are heavily covered by patents.


[10]  Why all the argument about file formats?

Strictly speaking, JPEG refers only to a family of compression algorithms;
it does *not* refer to a specific image file format.  The JPEG committee was
prevented from defining a file format by turf wars within the international
standards organizations.

Since we can't actually exchange images with anyone else unless we agree on
a common file format, this leaves us with a problem.  In the absence of
official standards, a number of JPEG program writers have just gone off to
"do their own thing", and as a result their programs aren't compatible with
anybody else's.

The closest thing we have to a de-facto standard JPEG format is some work
that's been coordinated by people at C-Cube Microsystems.  They have defined
two JPEG-based file formats:
  * JFIF (JPEG File Interchange Format), a "low-end" format that transports
    pixels and not much else.
  * TIFF/JPEG, aka TIFF 6.0, an extension of the Aldus TIFF format.  TIFF is
    a "high-end" format that will let you record just about everything you
    ever wanted to know about an image, and a lot more besides :-).  TIFF is
    a lot more complex than JFIF, and may well prove less transportable,
    because different vendors have historically implemented slightly different
    and incompatible subsets of TIFF.  It's not likely that adding JPEG to the
    mix will do anything to improve this situation.
Both of these formats were developed with input from all the major vendors
of JPEG-related products; it's reasonably likely that future commercial
products will adhere to one or both standards.

I believe that Usenet should adopt JFIF as the replacement for GIF in
picture postings.  JFIF is simpler than TIFF and is available now; the
TIFF 6.0 spec has only recently been officially adopted, and it is still
unusably vague on some crucial details.  Even when TIFF/JPEG is well
defined, the JFIF format is likely to be a widely supported "lowest common
denominator"; TIFF/JPEG files may never be as transportable.

A particular case that people may be interested in is Apple's QuickTime
software for the Macintosh.  QuickTime uses a JFIF-compatible format wrapped
inside the Mac-specific PICT structure.  Conversion between JFIF and
QuickTime JPEG is pretty straightforward, and several Mac programs are
available to do it (see Mac portion of section 6A).  If you have an editor
that handles binary files, you can strip a QuickTime JPEG PICT down to JFIF
by hand; see section 11 for details.

Another particular case is Handmade Software's programs (GIF2JPG/JPG2GIF and
Image Alchemy).  These programs are capable of reading and writing JFIF
format.  By default, though, they write a proprietary format developed by
HSI.  This format is NOT readable by any non-HSI programs and should not be
used for Usenet postings.  Use the -j switch to get JFIF output.  (This
applies to old versions of these programs; the current releases emit JFIF
format by default.  You still should be careful not to post HSI-format
files, unless you want to get flamed by people on non-PC platforms.)


[11]  How do I recognize which file format I have, and what do I do about it?

If you have an alleged JPEG file that your software won't read, it's likely
to be HSI format or some other proprietary JPEG-based format.  You can tell
what you have by inspecting the first few bytes of the file:

1.  A JFIF-standard file will start with the characters (hex) FF D8 FF E0,
    followed by two variable bytes (often hex 00 10), followed by 'JFIF'.

2.  If you see FF D8 at the start, but not the rest of it, you may have a
    "raw JPEG" file.  This is probably decodable as-is by JFIF software ---
    it's worth a try, anyway.

3.  HSI files start with 'hsi1'.  You're out of luck unless you have HSI
    software.  Portions of the file may look like plain JPEG data, but they
    won't decompress properly with non-HSI programs.

4.  A Macintosh PICT file, if JPEG-compressed, will have a couple hundred
    bytes of header followed by a JFIF header (scan for 'JFIF').  Strip off
    everything before the FF D8 and you should be able to read it.

5.  Anything else: it's a proprietary format, or not JPEG at all.  If you are
    lucky, the file may consist of a header and a raw JPEG data stream.
    If you can identify the start of the JPEG data stream (look for FF D8),
    try stripping off everything before that.

In uuencoded Usenet postings, the characteristic JFIF pattern is

	"begin" line
	M_]C_X ...

whereas uuencoded HSI files will start with

	"begin" line
	M:'-I ...

If you learn to check for the former, you can save yourself the trouble of
downloading non-JFIF files.


[12]  What about arithmetic coding?

The JPEG spec defines two different "back end" modules for the final output
of compressed data: either Huffman coding or arithmetic coding is allowed.
The choice has no impact on image quality, but arithmetic coding usually
produces a smaller compressed file.  On typical images, arithmetic coding
produces a file 5 or 10 percent smaller than Huffman coding.  (All the
file-size numbers previously cited are for Huffman coding.)

Unfortunately, the particular variant of arithmetic coding specified by the
JPEG standard is subject to patents owned by IBM, AT&T, and Mitsubishi.
Thus *you cannot legally use arithmetic coding* unless you obtain licenses
from these companies.  (The "fair use" doctrine allows people to implement
and test the algorithm, but actually storing any images with it is dubious
at best.)

At least in the short run, I recommend that people not worry about
arithmetic coding; the space savings isn't great enough to justify the
potential legal hassles.  In particular, arithmetic coding *should not*
be used for any images to be exchanged on Usenet.

There is some small chance that the legal situation may change in the
future.  Stay tuned for further details.


[13]  Does loss accumulate with repeated compression/decompression?

It would be nice if, having compressed an image with JPEG, you could
decompress it, manipulate it (crop off a border, say), and recompress it
without any further image degradation beyond what you lost initially.
Unfortunately THIS IS NOT THE CASE.  In general, recompressing an altered
image loses more information, though usually not as much as was lost the
first time around.

The next best thing would be that if you decompress an image and recompress
it *without changing it* then there is no further loss, i.e., you get an
identical JPEG file.  Even this is not true; at least, not with the current
free JPEG software.  It's essentially a problem of accumulation of roundoff
error.  If you repeatedly compress and decompress, the image will eventually
degrade to where you can see visible changes from the first-generation
output.  (It usually takes many such cycles to get visible change.)
One of the things on our to-do list is to see if accumulation of error can
be avoided or limited, but I am not optimistic about it.

In any case, the most that could possibly be guaranteed would be that
compressing the unmodified full-color output of djpeg, at the original
quality setting, would introduce no further loss.  Even such simple changes
as cropping off a border could cause further roundoff-error degradation.
(If you're wondering why, it's because the pixel-block boundaries move.
If you cropped off only multiples of 16 pixels, you might be safe, but
that's a mighty limited capability!)

The bottom line is that JPEG is a useful format for archival storage and
transmission of images, but you don't want to use it as an intermediate
format for sequences of image manipulation steps.  Use a lossless format
(PPM, RLE, TIFF, etc) while working on the image, then JPEG it when you are
ready to file it away.  Aside from avoiding degradation, you will save a lot
of compression/decompression time this way :-).


[14]  What are some rules of thumb for converting GIF images to JPEG?

As stated earlier, you *will* lose some amount of image information if you
convert an existing GIF image to JPEG.  If you can obtain the original
full-color data the GIF was made from, it's far better to make a JPEG from
that.  But if you need to save space and have only the GIF to work from,
here are some suggestions for getting maximum space savings with minimum
loss of quality.

The first rule when converting a GIF library is to look at each JPEG, to
make sure you are happy with it, before throwing away the corresponding GIF;
that will give you a chance to re-do the conversion with a higher quality
setting if necessary.  Some GIFs may be better left as GIFs, as explained in
section 3; in particular, cartoon-type GIFs with sixteen or fewer colors
don't convert well.  You may find that a JPEG file of reasonable quality
will be *larger* than the GIF.  (So check the sizes too.)

Experience to date suggests that large, high-visual-quality GIFs are the best
candidates for conversion to JPEG.  They chew up the most storage so offer
the most potential savings, and they convert to JPEG with least degradation.
Don't waste your time converting any GIF much under 100 Kbytes.  Also, don't
expect JPEG files converted from GIFs to be as small as those created
directly from full-color originals.  To maintain image quality you may have
to let the converted files be as much as twice as big as straight-through
JPEG files would be (i.e., shoot for 1/2 or 1/3rd the size of the GIF file,
not 1/4th as suggested in earlier comparisons).

Many people have developed an odd habit of putting a large constant-color
border around a GIF image.  While useless, this was nearly free in terms of
storage cost in GIF files.  It is NOT free in JPEG files, and the sharp
border boundary can create visible artifacts ("ghost" edges).  Do yourself
a favor and crop off any border before JPEGing.  (If you are on an X Windows
system, XV's manual and automatic cropping functions are a very painless
way to do this.)

cjpeg's default Q setting of 75 is appropriate for full-color input, but
for GIF inputs, Q settings of 85 to 95 often seem to be necessary to avoid
image degradation.  (If you apply smoothing as suggested below, the higher
Q setting may not be necessary.)

Color GIFs of photographs or complex artwork are usually "dithered" to fool
your eye into seeing more than the 256 colors that GIF can actually store.
If you enlarge the image, you will see that adjacent pixels are often of
significantly different colors; at normal size the eye averages these pixels
together to produce the illusion of an intermediate color value.  The
trouble with dithering is that, to JPEG, it looks like high-spatial-frequency
color noise; and JPEG can't compress noise very well.  The resulting JPEG
file is both larger and of lower image quality than what you would have
gotten from JPEGing the original full color image (if you had it).
To get around this, you want to "smooth" the GIF image before compression.
Smoothing averages together nearby pixels, thus approximating the color that
you thought you saw anyway, and in the process getting rid of the rapid
color changes that give JPEG trouble.  Appropriate use of smoothing will
often let you avoid using a high Q factor, thus further reducing the size of
the compressed file, while still obtaining a better-looking output image
than you'd get without smoothing.

With the V4 free JPEG software (or products based on it), a simple smoothing
capability is built in.  Try "-smooth 10" or so when converting GIFs.
Values of 10 to 25 seem to work well for high-quality GIFs.  Heavy-handed
dithering may require larger smoothing factors.  (If you can see regular
fine-scale patterns on the GIF image even without enlargement, then strong
smoothing is definitely called for.)  Too large a smoothing factor will blur
the output image, which you don't want.  If you are an image processing
wizard, you can also do smoothing with a separate filtering program, such as
pnmconvol from the PBMPLUS package.  However, cjpeg's built-in smoother is
a LOT faster than pnmconvol...

The upshot of all this is that "cjpeg -quality 85 -smooth 10" is probably a
good starting point for converting GIFs.  But if you really care about the
image, you'll want to check the results and maybe try a few other settings.


---------------------

For more information about JPEG in general or the free JPEG software in
particular, contact the Independent JPEG Group at [e-mail].

-- 
			tom lane
			organizer, Independent JPEG Group
Internet: [e-mail]	BITNET: tgl%cs.cmu.edu@carnegie