DV Pix - Multigeneration test images |
|
[updated 16 January 2000]
Multigeneration tests are useful for two purposes: they show off the strengths and limitations of the DV compression scheme, and they also show off any weak points in a particular implementation!
DVStressTest is my current test image. It's not perfect, but it does have a variety of test subjects:
The image DVtest1 is specifically designed to be hard to compress; it has "infinite" bandwidth with unnaturally sharp transitions horizontally, vertically, and diagonally. It has separate luma and chroma test sections, so that when you pass it through a DV codec you'll clearly see the difference between luma and chroma resolutions and compression methods. Each 16x16 square and its line of text is displaced one pixel horizontally and vertically in the DCT block structure from the preceding line, so each line will compress with different artifacts. Due to the sharp edges, multiple compression cycles (generations) will cause an increasing amount of mosquito noise to appear.
SF, on the other hand, is a naturalistic camera-captured image, which despite its high detail is already bandwidth limited by optical low-pass filtering in the camera and by one stage of DV compression. You can take it through multiple generations with little degradation occurring (or try your own camera-captured images).
Both images can be downloaded for your own tests. They are saved as TIFF files with LZW compression: this is a lossless compression that still yields acceptable sizes for transmission. To use them, simply import them into Premiere or another editing application; if your editor won't import TIFFs, use PhotoShop or another image editor to convert them to BMPs, PICTs, or another native format that your editing application can use. These images are 720x480 for 525/59.94 (NTSC) use; for use in 625/50 (PAL) format, open DVtest1 in Photoshop or another editor and add another 96 scanlines to make a 720x576 image (do not stretch the original image to fit the larger vertical space, or you'll destroy the relationship of the lines to the DCT block boundaries). Don't use "SF" for PAL tests, as its 4:1:1 structure won't be representative of the 4:2:0 images native to PAL DV formats (except for DVCpro, which uses 4:1:1 in PAL).
I suggest the following two tests. All instructions assume Adobe Premiere
4.2 / 5.X, but other editing applications offer similar operations, so
adjust the steps to suit your system. At each stage, I suggest saving the
resulting clip or still with an easily-remembered name followed by a
generation
number, i.e. "test01", "test02", "test03", etc., so that you can easily
keep track of all the stages of the test.
Using the Adaptec DVsoft codec in Premiere 4.2, I found that going ten
generations with DVtest1 adds a bit more mosquito noise in each
generation,
but the worst degradation happens with the first compression. Going ten
generations on SF showed no visible difference at all.
What I found using Premiere 4.2 and the Adaptec DVsoft codec on Windows
95 was that the luma information in this test was degraded exactly
as in the first test, as you would expect. Chroma, however,
underwent
a leftwards shift, drifting across the image with every new generation
-- and drifting across DCT block boundaries, too, causing further
degradation.
See the companion page
Pix: Codecs for some
5th-generation tests on another graphic image.
(c) 2000 Origin8 Video Note: JPEG compression shows artifacts not in the original |
This "pathological" text image is designed to stress the
DV
codec. Ramps show luma accuracy & linearity; bars show chroma
fidelity;
text & graphics show mosquito noise, chroma accuracy, Y/C
delay, etc.
Download DVStressTest.TIF (116 kBytes) |
Note: JPEG compression shows artifacts not in the original |
This "pathological" text image is designed
to stress
the DV codec; all the text is rendered with no antialiasing and
the sharpest
possible transitions.
The left side only has luma transitions; the right side has only chroma transitions; the luma value of the colors used is consistent (to a couple parts in a thousand) using the ITU-R BT601 equations converting RGB to YUV. On both sides, each line of text and its 16x16 pixel box is 31 pixels lower and one pixel to the right of the previous line. Thus each line is one pixel to the right and one pixel up, with respect to the DCT block structure*, compared to the previous line; each line will stress the DCT coding a bit differently. Download DVtest1.TIF (53 kBytes) *DV DCT luma blocks are 8x8 pixels for luma.
In 525/59.94,
the chroma blocks are 32 pixels wide and 8 tall; in 625/50 the
chroma blocks
are 16x16 (in both cases, the chroma blocks are 8x8 samples;
the odd sizes
are due to the 4:1:1 chroma subsampling in 525, and 4:2:0
subsampling in
625 systems)
|
Note: JPEG compression shows artifacts not in the original |
|
(c) 1995 Origin8 Video Note: JPEG compression shows artifacts not in the original |
This camera-captured image, shot with the 3-chip
DCR-VX1000
and exported from Adobe Premiere using the Adaptec DVsoft codec
under Windows
95, is more typical of the imagery found in video. While "sharp"
edges
are present, they are all prefiltered and bandwidth-limited by the
optical
and electronic processing of the video camera and circuitry -- and
it's
already been DV-compressed once.
This image shows little if any degradation in multigeneration tests: I've run it through the codec ten times, intercut 1st and 10th generation images in a looped clip, and been unable to see where the cuts occur. A mathematical comparison showed only minor differences between the two images, mostly around the edge of the building and the sun glare. Suggestion: instead of using this image, use one of your own. I only provide this one as a reference standard, so that when I say that I've done the test with a real image, that image is available for others to play with to verify my results. |
Copyright (c) 1998, 2000 by Adam J. Wilt.
You are granted a nonexclusive right to duplicate,
print,
link to, frame, or otherwise repurpose this material,
as long as all authorship, ownership and copyright
information
is preserved, and a link to this site is displayed.
Last updated 16 January 2000.