#image-resizing #resize #pixel-format #image-format #image

fast_image_resize

Library for fast image resizing with using of SIMD instructions

35 releases (15 stable)

3.0.4 Feb 15, 2024
2.7.3 May 7, 2023
2.7.0 Mar 24, 2023
2.4.0 Dec 11, 2022
0.1.0 Jul 31, 2021

#2 in Images

Download history 9048/week @ 2024-01-05 7917/week @ 2024-01-12 14345/week @ 2024-01-19 10293/week @ 2024-01-26 12190/week @ 2024-02-02 16488/week @ 2024-02-09 9136/week @ 2024-02-16 15034/week @ 2024-02-23 9992/week @ 2024-03-01 12601/week @ 2024-03-08 14440/week @ 2024-03-15 17505/week @ 2024-03-22 25975/week @ 2024-03-29 9986/week @ 2024-04-05 12124/week @ 2024-04-12 9261/week @ 2024-04-19

61,783 downloads per month
Used in 26 crates (20 directly)

MIT/Apache

730KB
16K SLoC

fast_image_resize

github crates.io docs.rs

Rust library for fast image resizing with using of SIMD instructions.

CHANGELOG

Supported pixel formats and available optimisations:

Format Description SSE4.1 AVX2 Neon Wasm32 SIMD128
U8 One u8 component per pixel (e.g. L) + + + +
U8x2 Two u8 components per pixel (e.g. LA) + + + +
U8x3 Three u8 components per pixel (e.g. RGB) + + + +
U8x4 Four u8 components per pixel (e.g. RGBA, RGBx, CMYK) + + + +
U16 One u16 components per pixel (e.g. L16) + + + +
U16x2 Two u16 components per pixel (e.g. LA16) + + + +
U16x3 Three u16 components per pixel (e.g. RGB16) + + + +
U16x4 Four u16 components per pixel (e.g. RGBA16, RGBx16, CMYK16) + + + +
I32 One i32 component per pixel - - - -
F32 One f32 component per pixel - - - -

Colorspace

Resizer from this crate does not convert image into linear colorspace during resize process. If it is important for you to resize images with a non-linear color space (e.g. sRGB) correctly, then you have to convert it to a linear color space before resizing and convert back to the color space of result image. Read more about resizing with respect to color space.

This crate provides the PixelComponentMapper structure that allows you to create colorspace converters for images whose pixels based on u8 and u16 components.

In addition, the crate contains functions create_gamma_22_mapper() and create_srgb_mapper() to create instance of PixelComponentMapper that converts images from sRGB or gamma 2.2 into linear colorspace and back.

Some benchmarks for x86_64

All benchmarks: x86_64, ARM64, WASM32.

Other libraries used to compare of resizing speed:

Resize RGB8 image (U8x3) 4928x3279 => 852x567

Pipeline:

src_image => resize => dst_image

  • Source image nasa-4928x3279.png
  • Numbers in table is mean duration of image resizing in milliseconds.
Nearest Box Bilinear Bicubic Lanczos3
image 28.20 - 82.45 134.07 192.70
resize - 26.83 53.56 97.73 144.63
libvips 7.73 60.66 19.84 30.15 39.46
fir rust 0.28 9.78 15.46 27.36 39.57
fir sse4.1 0.28 3.87 5.59 9.89 15.44
fir avx2 0.28 2.67 3.54 6.96 13.22

Resize RGBA8 image (U8x4) 4928x3279 => 852x567

Pipeline:

src_image => multiply by alpha => resize => divide by alpha => dst_image

  • Source image nasa-4928x3279-rgba.png
  • Numbers in table is mean duration of image resizing in milliseconds.
  • The image crate does not support multiplying and dividing by alpha channel.
Nearest Box Bilinear Bicubic Lanczos3
resize - 42.96 85.43 147.79 211.49
libvips 10.06 122.80 188.97 338.18 499.99
fir rust 0.19 20.10 27.08 41.32 56.79
fir sse4.1 0.19 10.03 12.24 18.57 25.15
fir avx2 0.19 6.98 8.26 13.97 21.55

Resize L8 image (U8) 4928x3279 => 852x567

Pipeline:

src_image => resize => dst_image

  • Source image nasa-4928x3279.png has converted into grayscale image with one byte per pixel.
  • Numbers in table is mean duration of image resizing in milliseconds.
Nearest Box Bilinear Bicubic Lanczos3
image 25.96 - 56.78 84.17 112.12
resize - 10.67 18.54 39.06 62.71
libvips 4.72 24.93 9.70 13.68 18.07
fir rust 0.15 4.08 5.24 7.48 11.33
fir sse4.1 0.15 1.86 2.30 3.58 5.88
fir avx2 0.15 1.66 1.86 2.24 4.21

Examples

Resize RGBA8 image

use std::io::BufWriter;
use std::num::NonZeroU32;

use image::codecs::png::PngEncoder;
use image::io::Reader as ImageReader;
use image::{ColorType, ImageEncoder};

use fast_image_resize as fr;

fn main() {
    // Read source image from file
    let img = ImageReader::open("./data/nasa-4928x3279.png")
        .unwrap()
        .decode()
        .unwrap();
    let width = NonZeroU32::new(img.width()).unwrap();
    let height = NonZeroU32::new(img.height()).unwrap();
    let mut src_image = fr::Image::from_vec_u8(
        width,
        height,
        img.to_rgba8().into_raw(),
        fr::PixelType::U8x4,
    ).unwrap();

    // Multiple RGB channels of source image by alpha channel 
    // (not required for the Nearest algorithm)
    let alpha_mul_div = fr::MulDiv::default();
    alpha_mul_div
        .multiply_alpha_inplace(&mut src_image.view_mut())
        .unwrap();

    // Create container for data of destination image
    let dst_width = NonZeroU32::new(1024).unwrap();
    let dst_height = NonZeroU32::new(768).unwrap();
    let mut dst_image = fr::Image::new(
        dst_width,
        dst_height,
        src_image.pixel_type(),
    );

    // Get mutable view of destination image data
    let mut dst_view = dst_image.view_mut();

    // Create Resizer instance and resize source image
    // into buffer of destination image
    let mut resizer = fr::Resizer::new(
        fr::ResizeAlg::Convolution(fr::FilterType::Lanczos3),
    );
    resizer.resize(&src_image.view(), &mut dst_view).unwrap();

    // Divide RGB channels of destination image by alpha
    alpha_mul_div.divide_alpha_inplace(&mut dst_view).unwrap();

    // Write destination image as PNG-file
    let mut result_buf = BufWriter::new(Vec::new());
    PngEncoder::new(&mut result_buf)
        .write_image(
            dst_image.buffer(),
            dst_width.get(),
            dst_height.get(),
            ColorType::Rgba8,
        )
        .unwrap();
}

Resize with cropping

use std::num::NonZeroU32;

use image::codecs::png::PngEncoder;
use image::io::Reader as ImageReader;
use image::{ColorType, GenericImageView};

use fast_image_resize as fr;

fn resize_image_with_cropping(
    mut src_view: fr::DynamicImageView,
    dst_width: NonZeroU32,
    dst_height: NonZeroU32
) -> fr::Image {
    // Set cropping parameters
    src_view.set_crop_box_to_fit_dst_size(
        dst_width, 
        dst_height, 
        None,
    );

    // Create container for data of destination image
    let mut dst_image = fr::Image::new(
        dst_width,
        dst_height,
        src_view.pixel_type(),
    );
    // Get mutable view of destination image data
    let mut dst_view = dst_image.view_mut();

    // Create Resizer instance and resize source image
    // into buffer of destination image
    let mut resizer = fr::Resizer::new(
        fr::ResizeAlg::Convolution(fr::FilterType::Lanczos3)
    );
    resizer.resize(&src_view, &mut dst_view).unwrap();

    dst_image
}

fn main() {
    let img = ImageReader::open("./data/nasa-4928x3279.png")
        .unwrap()
        .decode()
        .unwrap();
    let width = NonZeroU32::new(img.width()).unwrap();
    let height = NonZeroU32::new(img.height()).unwrap();
    let src_image = fr::Image::from_vec_u8(
        width,
        height,
        img.to_rgba8().into_raw(),
        fr::PixelType::U8x4,
    ).unwrap();
    resize_image_with_cropping(
        src_image.view(),
        NonZeroU32::new(1024).unwrap(),
        NonZeroU32::new(768).unwrap(),
    );
}

Change CPU extensions used by resizer

use fast_image_resize as fr;

fn main() {
    let mut resizer = fr::Resizer::new(
        fr::ResizeAlg::Convolution(fr::FilterType::Lanczos3),
    );
    unsafe {
        resizer.set_cpu_extensions(fr::CpuExtensions::Sse4_1);
    }
}

Dependencies

~0.4–1MB
~22K SLoC