DiffImageRotation.jl

This package serves only one purpose. To provide a bilinear interpolation or nearest neighbour based image rotation which works with CUDA and multithreaded CPUs (thanks to KernelAbstractions.jl). It can rotate images but also 3D arrays with a trailing batch dimension.

Further, it has registered adjoints with ChainRulesCore.jl.

Try it out with:

# install package with typing ]
# or ]add https://github.com/roflmaostc/DiffImageRotation.jl/
julia> ]add DiffImageRotation 

julia> using DiffImageRotation

julia> arr = zeros((32, 32)); arr[15:19, 10:23] .= 1 

julia> imrotate(arr, deg2rad(45));

julia> imrotate(arr, deg2rad(90));

julia> imrotate(arr, deg2rad(45), midpoint=(10,10))

julia> imrotate(arr, deg2rad(45), method=:nearest)
# access the docs
julia> ?imrotate

To learn more about the interpolation scheme, see this webpage. We implement rotation by area mapping (RAM).

Features

• bilinear interpolation based rotation
• nearest neighbor
• Automatic Differentation (AD) compatible
• KernelAbstractions.jl supported backends (CUDA, CPU, ...)
• midpoint around which rotation happens, can be specified

Related Packages

There is imrotate by ImageTransformations.jl. For standard image processing rather use this. It has better handling and way more options. But, it doesn't run with CuArrays and does not provide an adjoint/gradient rule.

There is rotate by FourierTools.jl. It's based on FFTs. It is based on a sinc interpolation.

Benchmarks

Tested on a AMD Ryzen 9 5900X 12-Core Processor with 24 Threads and a NVIDIA GeForce RTX 3060 with Julia 1.9.4 on Ubuntu 22.04.