This PR adds the modified zeng ("mzeng") palette sorting method, in
addition to the existing luma and battiato methods. Speed is very
similar to the battiato method with slightly better results on average.
Resulting sizes from two different image sets (all indexed or able to be
indexed):
| | master | PR |
|-|-|-|
| Set 1 | 29,647,156 | 29,555,697 |
| Set 2 | 23,732,133 | 23,570,862 |
Additionally, I've added a new "first colour" heuristic for both the
mzeng and battiato methods: We use the most popular colour overall, but
only if it covers at least 15% of the image. This provided 13k savings
on Set 2 vs the edge colour heuristic (which is still used in the luma
sort).
This adds a new palette sorting algorithm that attempts to minimise
entropy by an approximate solution to the Traveling Salesman Problem.
The algorithm comes from "An efficient Re-indexing algorithm for
color-mapped images" by Battiato et al
(https://ieeexplore.ieee.org/document/1344033).
It's fast and effective and works in addition to the luma sort (which
remains the single most effective sort). In order to keep lower presets
fast though, I've only enabled this for o3 and higher.
Results on a set of 190 indexed images at `-o5`:
18,932,727 bytes - master
18,578,306 bytes - PR
18,559,863 bytes - PR + #509
(These images may be particularly suited to alternative sorting methods
- the gains here are not necessarily what should be expected on average)
Note I looked into the 120 different palette sorting methods from
TruePNG, as mentioned in #74 (and seen in action in the Zopfli KrzYmod
fork). They're... largely ineffective. The combination of all 120
methods are outperformed by just the existing luma sort plus this new
one. That's not to say there's nothing further to be gained from them,
but trying to brute force all the combinations definitely seems like a
bad idea. There are other algorithms I hope to explore in future...
@ace-dent Thought this might interest you
UPDATE: I realised a quick tweak to alpha values in the luma sort can
provide a great improvement on images with transparency. The following
numbers were taken with PR #509 as base.
`-o2`:
19,065,549 bytes - base (luma sort)
18,949,747 bytes - modified luma sort
`-o5`:
18,922,165 bytes - base (luma sort)
18,559,863 bytes - new sorting algorithm + luma sort
18,544,813 bytes - new sorting algorithm + modified luma sort
* Keep track of number of pixels in each scanline
* Expand depth to 8-bit
* Attempt expand bit depth
* Simplify tracking of reduction_occurred
* Fix test
* Simplify depth handling in reductions
* Add tests for disabled reductions
* Fix bKGD conversion from gray to palette
* Allow grayscale reduction from 16 to 4 or less
* Refactor reduction evaluation sequence
* Separate palette into new file
* Refactor ColorType
Move transparency and palette data into the ColorType
* Fixup tests
* Make more use of helper functions
* Change BitDepth to u8 representation with TryFrom
* Fix clippy lints
* Don't use unstable language features
* Restore documentation on transparency/palette
* Refactor filters as enum
* Include filter byte in filter output
* Add entropy filter
* Add bigrams filter
* Add bigram entropy filter
* Add brute filter
* Replace bit-vec
* Add tests and benches
* Show filters in help
* Use FxHasher in color to palette
* Use windows function for minor improvement
For RGB(A) images that contain gray colors, this reduction can achieve
significant space savings. However, in the absence of gamma correction
data, some PNG decoders assume more exotic color spaces for grayscale
images instead of the ubiquitous sRGB. This results in gamma
miscorrection, and for the end user this means that colors will look
wrong, like "washed-out". Java's ImageIO class, which is popular in the
JVM world to read PNG files, uses rather unconventional defaults, as
explained in this StackOverflow question: https://stackoverflow.com/questions/31312645/java-imageio-grayscale-png-issue
Gamma miscorrection problems aside, OxiPNG currently tries hard to
reduce RGB(A) images to grayscale, because it expects that reduction to
be quite effective. However, in some cases, OxiPNG generates smaller
PNG files when reducing grasycale RGB(A) images to paletted color than
actual grayscale color. For example, let's say that "~/gray.png" is a
256x256 RGBA image entirely filled with (119, 119, 119, 255) pixels.
OxiPNG, by default, reduces this image to grayscale and achieves a
68.23% decrease:
$ cargo build --release && target/release/oxipng -omax --out ~/out.png ~/gray.png
Processing: /home/user/gray.png
256x256 pixels, PNG format
4x8 bits/pixel, RGBA
IDAT size = 604 bytes
File size = 661 bytes
Reducing image to 1x4 bits/pixel, Grayscale
Trying: 144 combinations
Found better combination:
zc = 6 zs = 0 f = 0 153 bytes
IDAT size = 153 bytes (451 bytes decrease)
file size = 210 bytes (451 bytes = 68.23% decrease)
Output: /home/user/out.png
However, if the --ng option that this commit adds is used to skip the
grayscale reduction step, OxiPNG reduces to a single color palette
instead, which is much more efficient, achieving a 84.42% decrease:
$ cargo build --release && target/release/oxipng -omax --ng --out ~/out.png ~/gray.png
Processing: /home/alejandro/gray.png
256x256 pixels, PNG format
4x8 bits/pixel, RGBA
IDAT size = 604 bytes
File size = 661 bytes
Reducing image to 1 bits/pixel, 1 colors in palette
Trying: 144 combinations
Found better combination:
zc = 3 zs = 3 f = 0 31 bytes
IDAT size = 31 bytes (573 bytes decrease)
file size = 103 bytes (558 bytes = 84.42% decrease)
Output: /home/alejandro/out.png
While OxiPNG should arguably be made smarter to better handle these
cases, in the meantime, adding an option to manually skip that grayscale
reduction can't hurt. In fact, it may even help users achieving the most
out of current versions of OxiPNG, and developers reasoning about what
makes a grayscale-like RGB(A) image compress better with a color
palette.
Due to the reasons stated above, this adds a simple "grayscale_reduction"
option to the Options struct, and a "no-grayscale-reduction" command
line switch, that makes OxiPNG not try this problematic grayscale
reduction on RGB(A) images.
* Improve performance consistency
Switch from HashMap / HashSet to IndexMap / IndexSet for consistent iteration order of various options and, as a result, more predictable performance.
