#convolution #image #computer-vision #kernel-module #image-processing #convolve

no-std convolve2d

Easy and extensible pure rust image convolutions

1 unstable release

0.1.0 Dec 9, 2021

#1317 in Algorithms

MIT license

45KB
524 lines

convolve2d: Image Convolutions in Rust

This crate defines an easy and extensible way to conduct image convolutions, in a way that is free of system dependencies, and works with no_std. Sound cool? Read on!

The purpose of convolve2d is to provide a single package that provides everything you need to conduct image convolutions suitable for computer vision or image manipulation. Here's a breif overview of what's on offer:

  • Two convolution functions: allowing you to pass your own buffer if speed is important, or have a buffer allocated and returned for a more idiomatic interface.

  • Traits: Convolution is defined generically across the Matrix trait. If you have a custom image type, simply define an implementation of Matrix, and you're good to go!

  • Built-in image Support: We also offer support for the image library through a feature flag (disabled by default), allowing you to seamlessly use the types you're already used to!

  • rayon: Compute convolutions in parallel using the rayon flag. (Enabled by default)

  • no_std Operation: to suit the needs of specialty systems or WASM.

  • Kernel Generators: The kernel module provides generation functions for a number of kernels commonly used in image processing.

While other convolution libraries may be more efficient, using a faster algorithm, or running on the GPU, this library's main focus is providing a complete convolution experience that is portable and easy to use.

Example:

This example shows how easy it is to perform convolutions when using the extensions for the image library. (See the image feature)

use image::RgbImage;
use convolve2d::*;

// Simply use `into` to convert from an `ImageBuffer` to a `DynamicMatrix`.
let image_buffer: RgbImage = ...;
let img: DynamicMatrix<SubPixels<u8, 3>> = image_buffer.into();

// Convert our color space to floating point, since our gaussian will be `f64`s
let img: DynamicMatrix<SubPixels<f64, 3>> = img.map_subpixels(|sp| sp as f64 / 255.0);

// Generate a 5x5 gaussian with standard deviation 2.0
let kernel = kernel::gaussian(5, 2.0);

// Perform the convolution, getting back a new `DynamicMatrix`
let convolution = convolve2d(&img, &kernel);

// Convert the color space back to 8-bit colors 
let convolution = convolution.map_subpixels(|sp| f64::round(sp * 255.0) as u8);

// Convert back into an `RgbImage` and save using the `image` library
RgbImage::from(convolution).save("output.png").expect("Unable to save image");

Features:

The following features are supported:

Feature Default Description
std Yes Allow access to the standard library, enabling the DynamicMatrix type.
rayon Yes Use rayon to compute convolutions in parallel.
image No Add extensions for interoperation with the image crate.
full No All features.

To use the library in no_std mode, simply disable all features:

convolve2d = { version = "0.1.0", default-features = false }

Notes on image Compatibility

Compatibility with the image library is provided using the image feature flag. This flag provides the following features:

  • The various pixel formats (Rgb, Luma, etc...) can now be converted to and from the SubPixels type. This allows them to be scaled and added as required for convolutions.
  • ImageBuffer can be converted to and from DynamicMatrixes using into and from.
  • ImageBuffers for which the pixel type is Luma can be used as Matrixes directly. This is because each element in the underlying data structure is one pixel. (Whereas in an RGB image, each element is one subpixel, meaning we need to group with SubPixels)

Acknowledgment:

Thanks to the following packages!

Crate Owner / Maintainer License
image HeroicKatora, fintelia MIT
rayon Josh Stone, Niko Matsakis Apache 2.0 or MIT
test-case Wojciech Polak, Luke Biel MIT

And to the Rust community at large!

Contributions:

Is something not clear in the documentation? Do we need another kernel type? This library came about as a personal project, but feel free to submit issues or PRs on GitLab!

License:

This crate is released under the terms of the MIT License.

Copyright (C) 2021 Joseph Skubal

Dependencies

~1.2–4MB
~36K SLoC