madler@piglet.caltech.edu (Mark Adler) (11/15/90)
>> The author is treating the "compression" like a dirty word. >> Compression need not decrease the quality of the image any. It depends >> on the type of compression. Yes it depends, and to get the compression ratios we're talking about, you certainly do not get back the original image bit-for-bit. The C-Cube chip has a user-selectable (approximate) compression ratio, and the higher you set that ratio, the worse the image gets on decompression. I have played with the "A" version of the chip and I find that ratios of 20 or so do not degrade the image very noticably. (Which is quite impressive if you think about it.) This assumes that the image is a "typical natural scene" without too much high-frequency stuff that a viewer might focus on (like text). The B version should be about the same---it simply conforms to the pretty much finished JPEG compression standard. JPEG compression uses every trick in the book (well not wavelets or fractals) to get those ratios. It breaks the image into 16x16 blocks and does a discrete-cosine-transform on them. This moves the high frequency stuff to the lower right corner of the box. Then the information in the box is quantized (this is where information is lost) and run-length encoded. This process is done separately on the Y, U, and V channels of the color image where Y is the luminance and U and V are the color-difference signals. U and V have half the resolution of the Y, so there is some more information lost there in the conversion from RGB. As a final note, JPEG is a standard for still images. There is MPEG in the works for moving pictures which compresses in the temporal direction and gets even higher ratios. It is similar to JPEG for the first frome, but subsequent frames (about 10 or so) are changes from the first frame with 16x16 blocks moving around. This scheme still has problems, and a tenatative standard may emerge in about a year. It will be longer before you see this in chips (there are some chips out now from LSI Logic, but they're just to speed up the experimentation). And there are other schemes, like wavelets and fractals, and who knows what else is around the corner ... Mark Adler madler@piglet.caltech.edu