- Notifications
You must be signed in to change notification settings - Fork14
NBLI is a fast, better lossless image compression algorithm. The optimized version (fNBLI) can get 1.5x compression ratio, 8x faster encoding, and 1.2x faster decoding compared to PNG. It can also beat the SOTA JPEG-XL on photographs.
License
WangXuan95/NBLI
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
NBLI (new-bee lossless image,v0.4) is a fast, better lossless compression algorithm, which support bothRGB 24-bit andGray 8-bit image. This repo provide:
- fNBLI : The optimized NBLI, currently only provide executable file without source code.
- NBLI : The prototype of NBLI, provide both executable file and source code.
Compared to the popular state-of-the-art lossless image formats,fNBLI is outstanding in both speed and compression ratio.
| Lossless Formats | Compressed Size (vs. PNG) | Compress Speed | Decompress Speed |
|---|---|---|---|
| BMF (-s) | -35.07% | 1.3 MB/s | 1.5 MB/s |
| Gralic | -34.32% | 4.2 MB/s | 3.5 MB/s |
| NBLI (-g) | -30.37% | 14.7 MB/s | 17.0 MB/s |
| JPEG-XL (-e 8) | -30.25% | 0.6 MB/s | 14.5 MB/s |
| fNBLI | -29.79% | 94.8 MB/s | 149.1 MB/s |
| JPEG-XL (-e 6) | -28.63% | 2.4 MB/s | 19.1 MB/s |
| JPEG-XL (-e 3) | -26.53% | 19.2 MB/s | 27.9 MB/s |
| WEBP (-5) | -22.53% | 3.7 MB/s | 82.8 MB/s |
| JPEG-XL (-e 1) | -13.49% | 95.2 MB/s | 71.0 MB/s |
| PNG (optipng) | 0.00% | 0.9 MB/s | 153.7 MB/s |
| JPEG2000 | +4.60% | 12.8 MB/s | 14.9 MB/s |
| PNG | +5.13% | 12.2 MB/s | 121.5 MB/s |
| Uncompressed | +114.77% |
The dataset of above comparison is thetraining image set ofCLIC2021 competition, which includes585 RGB 24-bit images, totaling4GB (the downloaded zip file is about 2GB). All above compressor/decompressor are tested in single-threaded, SeeLossless-Image-Compression-Benchmark for detail.
If you installed MinGW Compiler for Windows, you can run following command to compile NBLI, getting the executable fileNBLI.exe
g++ src\NBLI\main.cpp src\imageio\*.c src\imageio\uPNG\*.c src\NBLI\NBLI.cpp-static-O3-Wall-oNBLI.exe
Run following command to compile NBLI, getting the binary fileNBLI
g++ src/NBLI/main.cpp src/imageio/*.c src/imageio/uPNG/*.c src/NBLI/NBLI.cpp -static -O3 -Wall -o NBLI
RunNBLI.exe without any parameters to display its usage:
> .\NBLI.exe|----------------------------------------------------------------------------------|| NBLI: new-bee lossless image codec (single-threaded, v0.4, 202409)|| by https://github.com/WangXuan95/||----------------------------------------------------------------------------------|| Usage:|| NBLI [-switches]<in1> [-o<out1>] [<in2> [-o<out2]] ...|||| To compress:||<in> can be .pgm, .ppm, .pnm, or .png||<out> can be .nbli. It will be generatedif not specified.|||| To decompress:||<in> can be .nbli||<out> can be .pgm, .ppm, .pnm, or .png. It will be generatedif not specified.|||| switches: -v: verbose|| -f: force overwrite of output file|| -x: putting CRC32 when compressing|| -0~7: distortion level. 0=lossless (default), 1~7=lossy|| -g: use golomb arithmetic coding tree instead of ANS coding (slower)|| -a: use advanced predictor (extremely slow)||----------------------------------------------------------------------------------|
Take the images I provided in the./image/ folder as examples :
Compress a .png file to a .nbli file :
NBLI.exe -vf image\RGB.png -o RGB.nbliCompress a .png file to a .nbli file, use Golomb coding tree (-g) and Advanced predictor (-a) to obtain higher compression ratio :
NBLI.exe -vfga image\RGB.png -o RGB.nbliCompress a .png file to a .nbli file, set distortion level to 1 (lossy!!) :
NBLI.exe -vf1 image\RGB.png -o RGB.nbliDecompress the .nbli file back to a .png file :
NBLI.exe -vf RGB.nbli -o RGB.png
RunfNBLI.exe without any parameters to display its usage:
> .\fNBLI.exe|----------------------------------------------------------------------------------|| fNBLI: fast new-bee lossless image codec (single-threaded, v0.4, 202409)|| by https://github.com/WangXuan95/||----------------------------------------------------------------------------------|| this CPU: AVX2 supported||----------------------------------------------------------------------------------|| Usage:|| fNBLI [-switches]<in1> [-o<out1>] [<in2> [-o<out2]] ...|||| To compress:||<in> can be .pgm, .ppm, .pnm, or .png||<out> can be .fnbli. It will be generatedif not specified.|||| To decompress:||<in> can be .fnbli||<out> can be .pgm, .ppm, .pnm, or .png. It will be generatedif not specified.|||| switches: -v: verbose|| -f: force overwrite of output file|| -x: putting CRC32 when compressing||----------------------------------------------------------------------------------|
Take the images I provided in the./image/ folder as examples :
Compress a .pgm file (Gray image) to a .fnbli file :
fNBLI.exe -vf image\Gray.pgm -o Gray.fnbliCompress a .ppm file (RGB image) to a .fnbli file :
fNBLI.exe -vf image\RGB.ppm -o RGB.fnbliCompress a .png file to .fnbli file :
fNBLI.exe -vf image\RGB.png -o RGB.fnbliDecompress the .fnbli file back to a .ppm file :
fNBLI.exe -vf RGB.fnbli -o RGB.ppm
Decompress the .fnbli file back to a .png file :
fNBLI.exe -vf RGB.fnbli -o RGB.png
Thank the bottle of beer I drank on the evening after my PhD dissertation, which inspired me to come up with a novel parallel entropy encoding method that does not affect the compression ratio, which is the basis offNBLI.
As forNBLI, Its main method is not novel. The following works and peoples need to be mentioned:
- ANS-entropy-coding byJarosław Duda,Yann Collet, et al.
- CALIC: Context-based, adaptive, lossless image coding:https://ieeexplore.ieee.org/abstract/document/585919 by X. Wu et al.
- EDP: Edge-directed prediction for lossless compression of natural images:https://ieeexplore.ieee.org/abstract/document/923277 by Xin Li et al.
- Glicbawls: Gray Level Image Compression by Adaptive Weighted Least Squares:https://www.researchgate.net/publication/221578041_Glicbawls_-_Grey_Level_Image_Compression_by_Adaptive_Weighted_Least_Squares by Bernd Meyer et al.
- MRP: Lossless coding using variable block-size adaptive prediction optimized for each image:https://ieeexplore.ieee.org/document/7078076 by Ichiro Matsuda et al.
- A lossless color image compression method based on a new reversible color transform:https://ieeexplore.ieee.org/document/6410808 by Seyun Kim et al.
- JPEG-LS extension (ITU-T T.870)
- The experts of data compression,Alex Rhatushnyak,Matt Mahoney,Igor Pavlov, et al.
- And of course, Claude E. Shannon.
About
NBLI is a fast, better lossless image compression algorithm. The optimized version (fNBLI) can get 1.5x compression ratio, 8x faster encoding, and 1.2x faster decoding compared to PNG. It can also beat the SOTA JPEG-XL on photographs.