Image scaling library in Rust

Rust image scale in different color spaces using SIMD and multithreading.

Supported NEON, SSE, AVX-2, WASM.

Colorspace

This library provides for you some conveniences to scale in different color spaces.
Prebuilt options for CIE L*a*b, CIE L*u*v, CIE L*c*h, Linear, Sigmoidal, Oklab, Jzazbz available.
Those transformations also very efficients. Prefer downscale in linear colorspace or XYZ.
Up scaling might be done in LAB/LUB and simoidized components and also efficient in sRGB.

Have good f16 (binary float16) support.

Example integration with image crate

let img = ImageReader::open("./assets/asset.png")
    .unwrap()
    .decode()
    .unwrap();
let dimensions = img.dimensions();
let mut bytes = Vec::from(img.as_bytes());

let mut scaler = LinearScaler::new(ResamplingFunction::Lanczos3);
scaler.set_threading_policy(ThreadingPolicy::Adaptive);
// ImageStore::<u8, 4> - (u8, 4) represents RGBA, (u8, 3) - RGB etc
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, dimensions.0 as usize, dimensions.1 as usize).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(dimensions.0 as usize / 2, dimensions.1 as usize / 2),
    store,
    true
);
let resized_image = resized.as_bytes();

Fastest paths using SIMD

Despite all implementation are fast, not all the paths are implemented using SIMD, so some paths are slower

~ - Partially implemented

Features

For RISC-V riscv feature should be implicitly enabled, nightly compiler channel is required

To enable support of f16 the feature half should be activated.

Target features

neon optional target features are available, enable it when compiling on supported platform to get full features

avx2, fma, sse4.1, f16c will be detected automatically if available, and called the best path

fullfp16 NEON target detection performed in runtime, when available best the best paths for f16 images are available on ARM.

WASM simd128 target feature activating is mandatory in build flags

About f16

To enable full support of f16 half feature should be used, and f16c enabled when targeting x86 platforms. For NEON f16 feature use runtime detection, if CPU supports this feature then the very fast path is available

Even when half feature activated but platform do not support or features not enabled for f16 speed will be slow

Performance

Example comparison with fast-image-resize time for downscale RGB 4928x3279 image in 4 times.

Example comparison time for downscale RGBA 4928x3279 image in two times with premultiplying alpha.

Example comparison time for downscale RGBA 4928x3279 image in two times without premultiplying alpha.

Example comparison time for downscale RGBA 4928x3279 10 bit image in 2 times with premultiplying alpha.

RGBA 4928x3279 10 bit downscale 2 two times without premultiplying alpha

Example comparison time for downscale RGB 4000x6000 10 bit image in two times using NEON.

Example in sRGB

In common, you should not downsize an image in sRGB colorspace, however if speed is more preferable than more proper scale you may omit linearizing

let mut scaler = Scaler::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in linear

At the moment only sRGB transfer function is supported. This is also good optimized path so it is reasonably fast.

let mut scaler = LinearScaler::new(ResamplingFunction::Lanczos3);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in CIE L*a*b

let mut scaler = LabScaler::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in CIE L*u*v

let mut scaler = LuvScaler::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in CIE XYZ colorspace

let mut scaler = XYZScale::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in Sigmoidal colorspace

let mut scaler = SigmoidalScaler::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in LCh colorspace

let mut scaler = LChScaler::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Example in Oklab colorspace

let mut scaler = OklabScaler::new(ResamplingFunction::Hermite);
scaler.set_threading_policy(ThreadingPolicy::Single);
let store =
    ImageStore::<u8, 4>::from_slice(&mut bytes, width, height).unwrap();
let resized = scaler.resize_rgba(
    ImageSize::new(new_width, new_height),
    store,
    true
);

Resampling filters

Over 30 resampling filters is supported.

Bilinear
Nearest
Cubic
MitchellNetravalli
CatmullRom
Hermite
BSpline
Hann
Bicubic
Hamming
Hanning
Blackman

And others

This project is licensed under either of

• BSD-3-Clause License (see LICENSE)
• Apache License, Version 2.0 (see LICENSE)

at your option.