cykooz_resizer

cykooz_resizer is package with the optimized version of image resizing based on Rust's crate fast_image_resize.

CHANGELOG

Installation

python3 -m pip install cykooz-resizer

Or with automatically installing Pillow:

python3 -m pip install cykooz-resizer[pillow]

Information

Supported pixel types and available optimisations:

Implemented resize algorithms:

• Nearest - is nearest-neighbor interpolation, replacing every pixel with the nearest pixel in the output; for upscaling this means multiple pixels of the same color will be present.
• Convolution with different filters:
  • box
  • bilinear
  • catmull_rom
  • mitchell
  • gaussian
  • lanczos3
• Super sampling - is resizing an image in two steps. The first step uses the "nearest" algorithm. The second step uses "convolution" with a configurable filter.

Usage Examples

Resize Pillow's image

from PIL import Image

from cykooz_resizer import FilterType, ResizeAlg, Resizer, ResizeOptions


resizer = Resizer()
dst_size = (255, 170)
dst_image = Image.new('RGBA', dst_size)

for i in range(1, 10):
    image = Image.open('nasa_%d-4928x3279.png' % i)
    resizer.resize_pil(image, dst_image)
    dst_image.save('nasa_%d-255x170.png' % i)

# Resize using a bilinear filter and ignoring an alpha channel.
image = Image.open('nasa-4928x3279.png')
resizer.resize_pil(
    image,
    dst_image,
    ResizeOptions(
        resize_alg=ResizeAlg.convolution(FilterType.bilinear),
        use_alpha=False,
    )
)

Resize the raw image with an alpha channel

from cykooz_resizer import ImageData, PixelType, Resizer


def resize_raw(width: int, height: int, pixels: bytes):
    src_image = ImageData(
        width,
        height,
        PixelType.U8x4,
        pixels,
    )
    resizer = Resizer()
    dst_image = ImageData(255, 170, PixelType.U8x4)
    # By default, Resizer multiplies and divides by alpha channel
    # images with `U8x2`, `U8x4`, `U16x2` and `U16x4` pixels.
    resizer.resize(src_image, dst_image)
    return dst_image

Change used CPU-extensions

from cykooz_resizer import Resizer, CpuExtensions


resizer = Resizer()
resizer.cpu_extensions = CpuExtensions.sse4_1
...

Resize with using thread-pool

from cykooz_resizer import Resizer, ResizeOptions, ResizerThreadPool


...
thread_pool = ResizerThreadPool(num_threads=6)
resizer = Resizer()
resizer.resize(
    src_image,
    dst_image,
    ResizeOptions(thread_pool=thread_pool),
)
...

Benchmarks

Environment:

• CPU: AMD Ryzen 9 5950X
• RAM: DDR4 4000 MHz
• Ubuntu 24.04 (linux 7.0.0)
• Python 3.12
• Rust 1.97.1
• cykooz_resizer = "4.0" (single-threaded mode)

Other Python libraries used to compare of resizing speed:

• Pillow = "12.3.0" (https://pypi.org/project/Pillow/)

Resize algorithms:

• Nearest
• Convolution with Bilinear filter
• Convolution with Lanczos3 filter

Resize RGBA image 4928x3279 => 852x567

• Source image nasa-4928x3279.png

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

• Source image nasa-4928x3279.png has converted into a grayscale image with one byte per pixel.