real-time compression algorithms on fpga

[Melanie Nasic]

Hello community,

I am thinking about implementing a real-time compression scheme on an FPGA
working at about 500 Mhz. Facing the fact that there is no "universal
compression" algorithm that can compress data regardless of its structure
and statistics I assume compressing grayscale image data. The image data is
delivered line-wise, meaning that one horizontal line is processed, than
the next one, a.s.o.
Because of the high data rate I cannot spend much time on DFT or DCT and on
data modelling. What I am looking for is a way to compress the pixel data in
spatial not spectral domain because of latency aspects, processing
complexity, etc. Because of the sequential data transmission line by line a
block matching is also not possible in my opinion. The compression ratio is
not so important, factor 2:1 would be sufficient. What really matters is the
real time capability. The algorithm should be pipelineable and fast. The
memory requirements should not exceed 1 kb.
What "standard" compression schemes would you recommend? Are there
potentialities for a non-standard "own solution"?

Thank you for your comments.

Regards, Melanie

 

[DerekSimmons]

I attended Altera's DSP showcase/seminar in Rochester, there was a
large presence for interest in implementing MPEG4 (section 10 or 11, I
can't remember) for HDTV applications. You didn't say if the advice
you are searching for is for a commercial product, R&D science project
or a student project but even implementing something real time for DVD
quality (720x480) is worth considering.

I think a while back somebody announced the availability of JPEG
library on the www.opencores.org,
http://www.opencores.org/projects.cgi/web/jpeg/overview I haven't
tried it but critiquing it could be a good place to start. You could
incorporate parts of its implementation into yours, omit the parts you
don't agree with, and foster new not present in it. There is also a
section Video Compress Systems,
http://www.opencores.org/projects.cgi/web/video_systems/overview that
would be worth taking a look at.

A slightly different approach is using a tool like ImpulseC
(www.impulsec.com). It isn't really C but it is close. It allows you
to efficiently manage parallel systems of functions and integrate them
into VHDL and Verilog designs.

Maybe it is the bubble I have been living in, but your clock speed
seems high, what device families are you considering? I have seen 80
Mhz designs outperform similar applications on gigahertz PC. Don't
let PC marketing skew your objectivity (or maybe it is there choice in
operating systems).

Could you tell us more about the purpose of your project and you end
application?

Derek

 

[Stefan Rybacki]

Hello community,

...
What "standard" compression schemes would you recommend? Are there
potentialities for a non-standard "own solution"?

I would think of something like:
- run length encoding
- arithmetic coding
- maybe a dictionary encoding like zip
- if you know some statistics about the values you want to compress you could
also try if huffman coding is sufficent

Regards
Stefan

 

[=?iso-8859-1?q?Dag-Erling_Sm=F8rgrav?=]

Because of the high data rate I cannot spend much time on DFT or DCT
and on data modelling. What I am looking for is a way to compress
the pixel data in spatial not spectral domain because of latency
aspects, processing complexity, etc. Because of the sequential data
transmission line by line a block matching is also not possible in
my opinion. The compression ratio is not so important, factor 2:1
would be sufficient. What really matters is the real time
capability. The algorithm should be pipelineable and fast. The
memory requirements should not exceed 1 kb.

You don't say anything about quality.

Here's C code for a lossy compressor / decompressor which consistently
achieves a 2:1 ratio for 8 bpp grayscale images:

#include <stdint.h>
#include <stdio.h>

int
compress(FILE *fin, FILE *fout)
{
uint8_t pin[2], pout;

for (; {
if (fread(&pin, sizeof pin, 1, fin) != 1)
return (ferror(fin) ? -1 : 0);
pout = (pin[0] + pin[1]) / 2;
if (fwrite(&pout, sizeof pout, 1, fout) != 1)
return -1;
}
}

int
decompress(FILE *fin, FILE *fout)
{
uint8_t pin, pout[2];

for (; {
if (fread(&pin, sizeof pin, 1, fin) != 1)
return (ferror(fin) ? -1 : 0);
pout[0] = pout[1] = pin;
if (fwrite(&pout, sizeof pout, 1, fout) != 1)
return -1;
}
}

(note that the code assumes that the size of the input stream is an
even number)

DES

 

[Dave]

Hello community,

I am thinking about implementing a real-time compression scheme on an FPGA
working at about 500 Mhz. Facing the fact that there is no "universal
compression" algorithm that can compress data regardless of its structure
and statistics I assume compressing grayscale image data. The image data
is delivered line-wise, meaning that one horizontal line is processed,
than the next one, a.s.o.
Because of the high data rate I cannot spend much time on DFT or DCT and
on data modelling. What I am looking for is a way to compress the pixel
data in spatial not spectral domain because of latency aspects, processing
complexity, etc. Because of the sequential data transmission line by line
a block matching is also not possible in my opinion. The compression ratio
is not so important, factor 2:1 would be sufficient. What really matters
is the real time capability. The algorithm should be pipelineable and
fast. The memory requirements should not exceed 1 kb.
What "standard" compression schemes would you recommend? Are there
potentialities for a non-standard "own solution"?

Thank you for your comments.

Regards, Melanie

The answer, as always, is it all depends ...

Lossless compression (something like run length encoding) might work for
some kinds of image data (computer screens, rendered images), but will fail
for others (natural images etc).

Lossy compression will of course lose something from the image. The simplest
form is probably to average two adjacent pixels, giving you 2 to 1. I
suspect that anything more complex will exceed your space/speed/complexity
budget.

You need to spell out what type of images you are processing, what level of
'loss' is acceptable and why you need compression anyway (it will cause you
much pain !)

Dave

 

[Pete Fraser]

Hello community,

I am thinking about implementing a real-time compression scheme on an FPGA
working at about 500 Mhz. Facing the fact that there is no "universal
compression" algorithm that can compress data regardless of its structure
and statistics I assume compressing grayscale image data. The image data
is delivered line-wise, meaning that one horizontal line is processed,
than the next one, a.s.o.
Because of the high data rate I cannot spend much time on DFT or DCT and
on data modelling. What I am looking for is a way to compress the pixel
data in spatial not spectral domain because of latency aspects, processing
complexity, etc.

Are you hoping for lossless, or is lossy OK?

Because of the sequential data transmission line by line a block matching
is also not possible in my opinion. The compression ratio is not so
important, factor 2:1 would be sufficient. What really matters is the real
time capability. The algorithm should be pipelineable and fast. The memory
requirements should not exceed 1 kb.

That's 1024 bits, or bytes?
Is it enough for one line?
You don't say what your resolution and frame rate are.

What "standard" compression schemes would you recommend? Are there
potentialities for a non-standard "own solution"?

If you don't have a line of storage available, that restricts you a lot.
I don't understand this though. If you're using a 500MHz FPGA, it's
presumably recent, and presumably has a decent amount of storage.

How about a 1d predictor, non-linear quantizer and entropy coder?
If you have more memory available, look at JPEG-LS.
It can do lossless, or variable mild degrees of loss.

 

[Melanie Nasic]

Hi Pete,

I want the compression to be lossless and not based on perceptional
irrelevancy reductions. By stating 1 kb I meant 1024 bits and that's just
about half the line data. Your recommendation "1d predictor, non-linear
quantizer and entropy coder" sound interesting. COuld you please elaborate
on that? How is it done? Where can I find some exemplary codes? How can it
be achieved with hardware (VHDL sources?)

Thank you a lot.

Bye, Mel.

 

[Pete Fraser]

Hi Pete,

I want the compression to be lossless and not based on perceptional
irrelevancy reductions. By stating 1 kb I meant 1024 bits and that's just
about half the line data. Your recommendation "1d predictor, non-linear
quantizer and entropy coder" sound interesting. COuld you please elaborate
on that? How is it done? Where can I find some exemplary codes? How can it
be achieved with hardware (VHDL sources?)

I think you first need to play around with software and a few sample images.
The 1 d predictor means that you predict the next pixel in the sequence
by examining pixels on the left. A simple example would be to encode the
first pixel on the line, then use that as the prediction for the next pixel.
In that way you send only the difference between the predicted value
and what the pixel actually is. If you had enough memory to store a line you
could use a 2 d predictor, where you predict from pixels to the left and
pixels above.

Unfortunately, you can't use the non-linear quantizer as it's lossy.

I find Khalid Sayood's book "Introduction to Data Compression" quite good.
It comes with a link to a bunch of simple C code that has a variety of
predictors and entropy coders. You could try it on some sample images,
see how good compression you get, then go to hardware when you have
something acceptable.

 

[Brannon]

JPEGs are lossy because of the quantization step. You can do it without
the quantization step and still notice a significant compression. If
you preload your quantization constants and Huffman codes into lookup
tables, you can easily process one pixel per clock cycle in a 1500 gate
FPGA. I wrote a fully piplelined version that queued up the first eight
rows of an incoming image into block ram before starting on the DCTs.
It worked great. Ideally you would do DCTs on blocks larger than 8x8,
but the advantage to 8x8 is that you can easily do 64 8bit operations
in parallel which is nice for the Z-ordering, etc. Bigger squares
require bigger chips and external memory, and as soon as you have to go
to external memory you lose your pipelining.

You don't want to do a dictionary method for an image. In fact, I'm not
sure you want to do a dictionary method in FPGA. That sounds scary to
me. Use frequency domain (DCT or Wavelet) Z-ordering for photos and use
raw RLE for screen shots and similar images.

 

[Pete Fraser]

JPEGs are lossy because of the quantization step. You can do it without
the quantization step and still notice a significant compression. If
you preload your quantization constants and Huffman codes into lookup
tables, you can easily process one pixel per clock cycle in a 1500 gate
FPGA. I wrote a fully piplelined version that queued up the first eight
rows of an incoming image into block ram before starting on the DCTs.
It worked great. Ideally you would do DCTs on blocks larger than 8x8,
but the advantage to 8x8 is that you can easily do 64 8bit operations
in parallel which is nice for the Z-ordering, etc. Bigger squares
require bigger chips and external memory, and as soon as you have to go
to external memory you lose your pipelining.

Is the OP not getting pixels in raster-scan order though?
That, and a half-line memory limit means there's not enough storage
for a 2-d DCT.

 

[Nils]

If you can store a buffer of at least one extra scanline, you could try the
Paeth predictor + RLE. This will give reasonable prediction of the next
pixel's grayscale value, and if the prediction is OK, the result will often
contain a string of zeroes, and the RLE will do a good job.

If you can "afford it" (in other words, the FPGA is fast enough), you could
use arithmetic coding on the resulting predicted values, with a simple order
0 model, instead of RLE.

Paeth + RLE will do OK on computer generated images, but not on natural
images. Paeth + AC will do OK on both.

Both will fit in 1kb of code for sure.

Nils

 

[John B]

Hello community,

I am thinking about implementing a real-time compression scheme on an
FPGA working at about 500 Mhz. Facing the fact that there is no
"universal compression" algorithm that can compress data regardless
of its structure and statistics I assume compressing grayscale image
data. The image data is delivered line-wise, meaning that one
horizontal line is processed, than the next one, a.s.o. Because of
the high data rate I cannot spend much time on DFT or DCT and on data
modelling. What I am looking for is a way to compress the pixel data
in spatial not spectral domain because of latency aspects, processing
complexity, etc. Because of the sequential data transmission line by
line a block matching is also not possible in my opinion. The
compression ratio is not so important, factor 2:1 would be
sufficient. What really matters is the real time capability. The
algorithm should be pipelineable and fast. The memory requirements
should not exceed 1 kb. What "standard" compression schemes would
you recommend? Are there potentialities for a non-standard "own
solution"?

Thank you for your comments.

Regards, Melanie

Have a look at Graphics File Formats by Kay & Levine (ISBN
0-07-034025-0). It will give you some ideas.

 

[news]

: I want the compression to be lossless and not based on perceptional
: irrelevancy reductions.

If it has to be lossless there's no way you can guarantee to
get 2:1 compression (or indeed any compression at all!). You
may do, with certain kinds of input, but it's all down to the
statistics of the data. The smaller your storage the less
you can benefit from statistical variation across the image,
and 1 Kbyte is very small!

Given that a lossless system is inevitably 'variable bit rate'
(VBR) the concept of "real time capability" is somewhat vague;
the latency is bound to be variable. In real-world applications
the output bit-rate is often constrained so a guaranteed minimum
degree of compression must be achieved; such systems cannot be
(always) lossless.

From my experience I would say you will need at least a 4-line
buffer to get near to 2:1 compression on a wide range of input
material. For a constant-bit-rate (CBR) system based on a 4x4
integer transform see:

http://www.bbc.co.uk/rd/pubs/whp/whp119.shtml

This is designed for ease of hardware implementation rather than
ultimate performance, and is necessarily lossy.

Richard.
http://www.rtrussell.co.uk/
To reply by email change 'news' to my forename.

 

[Jerry Coffin]

(e-mail address removed) wrote:

[ ... ]

Given that a lossless system is inevitably 'variable bit rate'
(VBR) the concept of "real time capability" is somewhat vague;
the latency is bound to be variable.

The output bit rate will vary, but can be bounded -- for an obvious
example, consider using LZW compression with 8-bit inputs and 12-bit
codes as the output. In the worst case, each 8-bit input produces a
12-bit output, and you have to clear and rebuild the dictionary every
4096 characters.

Real-time constraints are a more or less separate issue though -- here,
the output data stream isn't (usually) nearly as difficult to deal with
as things like dictionary searching in the compression process. Like
the compression rate, this will vary, but (again) it's usually pretty
easy to set an upper bound. Using the same example, in LZW you build up
a string out of characters, with one dictionary entry for each "next"
character following a current character. There can be no more than 256
next characters (worst case), so the worst case requirement is to
search all 256 entries for follow-on characters in N microseconds (or
nanoseconds, or whatever). You just about need more details to
guarantee this though -- a trie-based dictionary has different
characteristics than a hash-based dictionary (for an obvious example).
In nearly every case it's still pretty easy to place an upper bound on
the complexity and time involved though.

OTOH, you can run into a little bit of a problem with some of the
basics -- if you happen (for example) to be storing your dictionary in
SRAM, the time per access is pretty easy to estimate. If you're storing
it in something like SDRAM, the worst case can be a bit harder to
figure out.

 

[Jerry Coffin]

Hello community,

I am thinking about implementing a real-time compression scheme on an FPGA
working at about 500 Mhz. Facing the fact that there is no "universal
compression" algorithm that can compress data regardless of its structure
and statistics I assume compressing grayscale image data. The image data is
delivered line-wise, meaning that one horizontal line is processed, than
the next one, a.s.o.
Because of the high data rate I cannot spend much time on DFT or DCT and on
data modelling. What I am looking for is a way to compress the pixel data in
spatial not spectral domain because of latency aspects, processing
complexity, etc. Because of the sequential data transmission line by line a
block matching is also not possible in my opinion. The compression ratio is
not so important, factor 2:1 would be sufficient. What really matters is the
real time capability. The algorithm should be pipelineable and fast. The
memory requirements should not exceed 1 kb.
What "standard" compression schemes would you recommend?

JPEG supports lossless encoding that can fit (at least roughly) within
the constraints you've imposed. It uses linear prediction of the
current pixel based on one or more previous pixels. The difference
between the prediction and the actual value is what's then encoded. The
difference is encoded in two parts: the number of bits needed for the
difference and the difference itself. The number of bits is Huffman
encoded, but the remainder is not.

This has a number of advantages. First and foremost, it can be done
based on only the curent scan line or (depending on the predictor you
choose) only one scan line plus one pixel. In the latter case, you need
to (minutely) modify the model you've outlined though -- instead of
reading, compressing, and discarding an entire scan line, then starting
the next, you always retain one scan line worth of data. As you process
pixel X of scan line Y, you're storing pixels 0 through X+1 of the
current scan line plus pixels X-1 through N (=line width) of the
previous scan line.

Another nice point is that the math involved is always simple -- the
most complex case is one addition, one subtraction and a one-bit right
shift.

Are there
potentialities for a non-standard "own solution"?

Yes, almost certainly. Lossless JPEG is open to considerable
improvement. Just for an obvious example, it's pretty easy to predict
the current pixel based on five neighboring pixels instead of three. At
least in theory, this should improve prediction accuracy by close to
40% -- thus reducing the number of bits needed to encode the difference
between the predicted and actual values. At a guess, you won't really
see 40% improvement, but you'll still see a little improvement.

In the JPEG 2000 standard, they added JPEG LS, which is certainly an
improvement, but if memory serves, it requires storing roughly two full
scan lines instead of roughly one scan line. OTOH, it would be pretty
easy to steal some of the ideas in JPEG LS without using the parts that
require more storage -- some things like its handling of runs are
mostly a matter of encoding that shouldn't really require much extra
storage.

The final question, however, is whether any of these is likely to give
you 2:1 compression. That'll depend in your input data -- for typical
photographs, I doubt that'll happen most of the time. For thngs like
line art, faxes, etc., you can probably do quite a bit better than 2:1
on a fairly regular basis. If you're willing to settle for nearl
lossless compression, you can improve ratios a bit further.

 

[Jerry Coffin]

Hello community,

I am thinking about implementing a real-time compression scheme on an FPGA
working at about 500 Mhz. Facing the fact that there is no "universal
compression" algorithm that can compress data regardless of its structure
and statistics I assume compressing grayscale image data. The image data is
delivered line-wise, meaning that one horizontal line is processed, than
the next one, a.s.o.
Because of the high data rate I cannot spend much time on DFT or DCT and on
data modelling. What I am looking for is a way to compress the pixel data in
spatial not spectral domain because of latency aspects, processing
complexity, etc. Because of the sequential data transmission line by line a
block matching is also not possible in my opinion. The compression ratio is
not so important, factor 2:1 would be sufficient. What really matters is the
real time capability. The algorithm should be pipelineable and fast. The
memory requirements should not exceed 1 kb.
What "standard" compression schemes would you recommend?

Though it's only rarely used, there's a lossless version of JPEG
encoding. It's almost completely different from normal JPEG encoding.
This can be done within your constraints, but would be improved if you
can relax them minutely. Instead of only ever using the current scan
line, you can improve things if you're willing to place the limit at
only ever storing one scan line. The difference is that when you're in
the middle of a scan line (for example) you're storing the second half
of the previous scan line, and the first half of the current scan line,
rather than having half of the buffer sitting empty. If you're storing
the data in normal RAM, this makes little real difference -- the data
from the previous scan line will remain in memory until you overwrite
it, so it's only really a question of whether you use it or ignore it.

Are there potentialities for a non-standard "own solution"?

Yes. In the JPEG 2000 standard, they added JPEG LS, which is another
lossless encoder. A full-blown JPEG LS encoder needs to store roughly
two full scan lines if memory serves, which is outside your
constraints. Nonetheless, if you're not worried about following the
standard, you could create more or less a hybrid between lossless JPEG
and JPEG LS, that would incorporate some advantages of the latter
without the increased storage requirements.

I suspect you could improve the prediction a bit as well. In essence,
you're creating a (rather crude) low-pass filter by averaging a number
of pixels together. That's equivalent to a FIR with all the
coefficients set to one. I haven't put it to the test, but I'd guess
that by turning it into a full-blown FIR with carefully selected
coefficients (and possibly using more of the data you have in the
buffer anyway) you could probably improve the predictions. Better
predictions mean smaller errors, and tighter compression.

 

[Melanie Nasic]

Hi Jerry,

thanks for your response(s). Sounds quite promising. Do you know something
about hardware implementation of the compression schemes you propose? Are
there already VHDL examples available or at least C reference models?

Regards, Melanie

 

[Jerry Coffin]

Hi Jerry,

thanks for your response(s).

Sorry 'bout that -- Google claimed the first one hadn't posted, so I
tried again...

Sounds quite promising. Do you know something
about hardware implementation of the compression schemes you propose? Are
there already VHDL examples available or at least C reference models?

I don't believe I've seen any VHDL code for it. One place that has C
code is:

ftp://ftp.cs.cornell.edu/pub/multimed/ljpg.tar.Z

I suspect Google would turn up a few more implementations as well. If
you like printed information, you might consider _Image and Video
Compression Standards: Algorithms and Architectures; Second Edition_ by
Bhaskaran and Konstantinides. Published by Kluwer Academic Publishers,
ISBN 0-7923-9952-8. It doesn't go into tremendous detail, but it's one
of the few (that I know of) that discusses lossless JPEG at all.

 

[=?ISO-8859-1?Q?Michael_Sch=F6berl?=]

I am thinking about implementing a real-time compression scheme on an FPGA

working at about 500 Mhz.

I'm currently working on something similar ...



simple predictive schemes (like the 4th predictor from jpeg or MED (from
jpeg-ls)) look promising ... they require storage of one line

entropy coding:
huffman and multilevel arithmetic coding require a lot of ressources
(that easily ends up above 1kbit even for a table of probabilities)
binary arithmetic coding would be able to code less than 1bit/cycle
(which ends up at >5 cyles/pixel which is too slow in my case)

I'm currently investigating different schemes of golomb-rice codes
(static, adaptive or even context-adaptive like in jpeg-ls) ... so far
they look promising ...


jpeg-ls: the concept looks nice and quite powerful for a pipelined
encoder (like for fpgas) - unfortunately the decoder would require a
large feedback-loop (and pipelining is almost impossible) ...
jpeg-ls is only symmetric for software implementations :-(



I'm still curious how you plan to achieve 500MHz even on a Virtex4
(I would say something like 200MHz could be possible)


bye,
Michael

 

[Thomas Rudloff]

Hi Pete,

I want the compression to be lossless and not based on perceptional
irrelevancy reductions. By stating 1 kb I meant 1024 bits and that's just
about half the line data. Your recommendation "1d predictor, non-linear
quantizer and entropy coder" sound interesting. COuld you please elaborate
on that? How is it done? Where can I find some exemplary codes? How can it
be achieved with hardware (VHDL sources?)

Thank you a lot.

Bye, Mel.

Hi Mel,

you can calculate the delta of two subsequent pixels an then Huffman
code this result. This should archive almost 2:1 if there are not much
large brightness steps in the picture.

Regards
Thomas