Close to-Lossless Picture Codecs Utilizing Extremely-Quick LZ Codecs
I believe easy ultra-high velocity lossy (or near-lossless) picture codecs, constructed from the brand new era of quick LZ codecs, are going to change into extra related sooner or later.
Computing bottlenecks change over time. As disk area, disk bandwidth, and web bandwidth will increase, older picture codecs that squeeze each final bit out of the ensuing file change into much less beneficial for a lot of use instances. Finally web sites or merchandise utilizing noticeably lossy compressed photos can be much less enticing. The bandwidth financial savings from overly lossy picture codecs will change into meaningless, and the CPU/consumer time and battery or grid power spent on advanced decompression steps can be wasted.
Finally, a lot easier codecs with decrease (weaker) compression ratios that introduce much less distortion, however have blazingly quick decompression charges are going to change into extra widespread. This core idea motivates this work.
One technique to assemble a easy lossless or lossy picture codec with quick decompression is to mix a customized encoding software with the favored lossless LZ4 codec. LZ4’s compression and decompression is extraordinarily quick, and the library is dependable, up to date usually, extensively fuzz examined, and quite simple to make use of.
To make it lossy, the encoder must precondition the picture information so when it is subsequently compressed by LZ4, the proportion of 4+ byte matches vs. literals is elevated in comparison with the unique picture information. I have been developing LZ Preconditioners, and constructing new LZ codecs that amend themselves to this preconditioning step, over the previous 12 months.
Such a codec will be unable to compete in opposition to JPEG, WebP, JPEG 2000, and so on. for perceived high quality per bit. Nonetheless, it will be extraordinarily quick to decode, quite simple, and can possible not bloat executables as a result of the LZ4 library is already current in lots of codebases. Utilizing LZ4 introduces no new safety dangers.
This LZ preconditioning step should be accomplished in a approach that minimizes apparent seen artifacts, as perceived by the Human Visual System (HVS). This tradeoff, of accelerating distortion however lowering the bitrate is a traditional software of Rate-distortion theory. That is well-known in video coding, and now in GPU texture encoding (which I launched in 2009 with my “crunch” compression library).
The rdopng tool on github helps creating lossy LZ4 compressed RGB/RGBA photos utilizing a easy rate-distortion mannequin. (Lossless is subsequent, however I needed to unravel the tougher drawback first.) In the course of the preconditioning step, the LZ4 charge in bits is approximated utilizing a sliding dictionary and a match finder. For every potential match substitute which might introduce distortion into the lookahead buffer, the preconditioner approximates the launched visible error by computing colour error distances in a scaled Oklab perceptual colorspace. (Oklab is likely one of the strongest colorspaces I’ve used for this kind of work. There are higher colorspaces for compression, however Oklab is easy to make use of and well-documented.)
Perceptually, distortions launched into areas of photos surrounded by excessive frequency particulars are much less noticeable vs. areas containing easy or gradient options. Earlier than preconditioning, the encoder computes two error scaling masks which point out which areas of the picture comprises massive or small gradients/easy areas. These scaling masks suppress introducing distortions (through the use of longer or extra matches) if doing so can be too noticeable to the HVS. This step has a big influence on bitrate and might be improved.
To hurry up encoding, the preconditioner solely examines a window area above and to the left of the lookahead buffer. LZ4’s unlucky minimal match measurement of 4 bytes complicates encoding of 24bpp RGB photos. Encoding just isn’t very quick resulting from this search course of, however it’s doable to string it by engaged on completely different areas of the picture in parallel. The encoder is a proof of precept and testing grounds, and never as quick because it may very well be, however it works.
The encoder additionally helps angular error metrics for encoding tangent area regular maps.
LZ4I photos are trivial to decode in any language. A LZ4I picture consists of a easy header adopted by the LZ4 compressed 24bpp RGB (R first) or RGBA pixel information:
Some instance encodings:
Lossy LZ4I: 42.630 709 Y dB, 12.985 bits/pixel, 1.1 gigapixels/sec. decompression charge utilizing LZ4_decompress_safe() (on a cell Core i7 1065G7 at 1.3GHz base clock):
Biased delta picture:
Greyscale histogram of biased delta picture:
A extra lossy LZ4I encoding, 38.019 709 Y dB, 8.319 bits/pixel:
Biased delta picture:
Greyscale delta picture histogram:
Lossless unique: 14.779 bits/pixel PNG, 16.551 bits/pixel LZ4I:
Lossy LZ4I: 45.758 709 Y dB, 7.433 bits/pixel (lower than half the scale vs. lossless LZ4I!):
Biased delta picture:
LZSSE is especially enticing as a preconditioning goal as a result of it is 30-40% sooner than LZ4 and has the next ratio. That is subsequent. A format that makes use of tiled decompression and a number of threads can be a no brainer. Finally I believe LZ or QOI-like variants can be very sturdy contenders sooner or later.