As a caveat before beginning, let me note that I am by no means a compression expert at the algorithm level. Nevertheless, if you are a practicing video/multimedia producer, and have a slightly technical bent, I think you will find the below of significant interest.
David Durlach
December 5th, 2007
A clear example of the issue
In the image below, there are four rows, each labeled as to its contents.
[You may immediately object to the fact that the top row is labeled "No Compression", given that the web image itself is JPEG compressed. However, it is visually indistinguishable from the original BMP file, having been compressed with Photoshop CS2's Maximum ("12") compression setting. Furthermore, a 4.4 meg zip file containing the truly uncompressed source image, along with many more foreground/background text color combinations than are shown below, and the JPEG compression artifacts evidenced therein, may be downloaded here. I strongly encourage so downloading.
Also, note that the second column photo, coming from a relatively low quality digital camera, is noticeably JPEG "smeared" even in the "Uncompressed" version. That photo should thus be substantially ignored in terms of the discussion below. (In retrospect I should have tried to find a better shot, but as its presence is sort of random anyway, I did not think it worth redoing everything just for that.)]
You will notice that rows 2 and 3 above are essentially identical in file size, each being JPEG compressed by about 10-1 compared to the original uncompressed source. Yet in row 2, the blue text on a black background looks horrible, while in row 3 it looks fine and is essentially indistinguishable from the original uncompressed version in row 1. Why is this given the files sizes are the same? The answer is that in row 2, color subsampling was performed (I believe 4:2:2), while in row 3, to the best of my knowledge, there was no subsampling at all (i.e., "4:4:4"). (It is also interesting to note that the color subsampling has no apparent detrimental affect on the B&W text/background.)
Finally, just for reference, row 4 shows extreme compression (way too much for most projects) of over 50-1. Here the "standard" problems of JPEG compression show up, including color "smearing" even in the black & white text, and the photo for the first time shows noticeable degradation (even beyond its original smearing). Note that this row was generated by the lowest possible JPEG quality setting (0) in Adobe After Effects (AE) V7.0, exporting a "JPEG Sequence".
To summarize, if you see problems with text degradation in JPEG compressed images, unless you have compressed a HUGE amount, the problem is almost certainly not intrinsic to the JPEG codec, but rather to a less-than-ideal color subsampling scheme. Much additional information, including what common codecs perform color subsampling, follows.
MJPEG, or Motion JPEG, as the reader may know, is simply a data structure wrapping individual frames which are each themselves JPEG-compressed. (Do not confuse MJPEG with MPEG). Thus, there are (for example) both Quicktime and AVI movies whose codec (Compression/Decompression algorithm) is MJPEG.
MJPEG/JPEG is only loosely defined as a compression scheme - there is much variability in how JPEG frame compression is actually implemented in any particular case. Additionally, the single "Quality" setting in many application's MJPEG/JPEG compression dialog box, be it a percentage value 0-100, or a "Low", "Medium", "High", "Best" choice, is not well defined in terms of what effect it will have on (each step of) the (loosely-defined) algorithm. This state of affairs has very real-world impact as we have already seen above.
I focus on MJPEG (as opposed to simply JPEG) for several reasons. First, the creators of popular MJPEG compressors have often made different algorithm choices than those creating popular single frame JPEG compressors. Second, Quicktime MJPEG files are what TechnoFrolics' Spin Browser VE requires. Finally, file size when rendering hours of video is frequently a lot more important that when creating just a few dozen still images.
Finally, note that the factor by which one can compress an image without visible artifacts will vary greatly from image to image.
MJPEG/JPEG compression has a number of relatively separate steps to the algorithm. More detail than you probably want can be found in Wikipedia here, with further discussion of color space conversion and sub-sampling in the "Fundamentals of embedded video, Part 2" article written by Analog Devices engineers. A much simpler and less mathematical page, with images, and duplicating some of the information presented in our discussion (page was discovered after most here was already written), but focused on JPEG (as opposed to MJPEG), can be found at Optimization of JPEG compression settings. The concepts that must be understood for our discussion, described in much more detail in the above first two links, are:
Phase 1: Perform (or not), color subsampling. This is where adjacent pixels "share" a single color designator in order to reduce file size. There are two key points to understand about this: First and most important, this is is a discrete image-degrading step - in other words, there is (typically anyway) no "slider" to continuously adjust the quality loss resulting from this step. Second, that the exact algorithm for such color subsampling is not necessarily well defined (though for some JPEG standards it may be).
Phase 2: Perform an operation similar to a two dimensional Fourier transform, and then throw away a certain fraction of the high frequency components. This step is where the image is converted into "high and low frequency components", where high frequencies correspond to sharp color transitions and fine detail, and low frequency components correspond to the larger, softer scene elements. What is key to understand here is that the typical "Quality Slider" in an application's MJPEG/JPEG compression dialog box determines, in a continuously adjustable way, how much of the high frequency info is discarded. (However, the quality slider may also affect the color subsampling scheme!)
(END)
