Julia interface to libjpeg-turbo
Author JuliaIO
13 Stars
Updated Last
1 Year Ago
Started In
January 2022


Stable Dev Build Status Coverage

JpegTurbo.jl is a Julia wrapper of the C library libjpeg-turbo that provides IO support for the JPEG image format. This package also backs the JPEG IO part of ImageIO and FileIO.

For benchmark results against other image IO backends, please check here.


There are two different usages for this package:

  • (convenient) via the FileIO: save/load
  • (powerful) via the JpegTurbo.jl interfaces: jpeg_encode/jpeg_decode

FileIO interface: save/load

FileIO is an IO frontend with various IO backends; ImageIO is the default IO backend provided by the JuliaImages ecosystem. When JpegTurbo (and/or ImageIO) are available in DEPOT_PATH, FileIO will uses JpegTurbo to load and save the JPEG images:

using FileIO
img = rand(64, 64)
save("test.jpg", img)

Note that you do not necessarily need to install them in your project environments. For instance, you can do (@v1.8) pkg> add JpegTurbo or (@v1.8) pkg> add ImageIO and it should work for your local setup.

JpegTurbo interface: jpeg_encode/jpeg_decode

jpeg_encode is used to compress 2D colorant matrix as JPEG image.

jpeg_encode(filename::AbstractString, img; kwargs...) -> Int
jpeg_encode(io::IO, img; kwargs...) -> Int
jpeg_encode(img; kwargs...) -> Vector{UInt8}

jpeg_decode is used to decompress JPEG image as 2D colorant matrix.

jpeg_decode([T,] filename::AbstractString; kwargs...) -> Matrix{T}
jpeg_decode([T,] io::IO; kwargs...) -> Matrix{T}
jpeg_decode([T,] data::Vector{UInt8}; kwargs...) -> Matrix{T}

Advanced: in-memory encode/decode

For some applications, it can be faster to do encoding/decoding without the need to read/write disk:

using JpegTurbo
img = rand(64, 64)
bytes = jpeg_encode(img) # Vector{UInt8}
img_saveload = jpeg_decode(bytes) # size: 64x64

Advanced: preview optimization

One can request a single-component output or a downsampled output so that fewer calculation is needed during the decompression. This can be particularly useful to accelerate image preview.

using BenchmarkTools, TestImages, JpegTurbo
filename = testimage("earth", download_only=true)
# full decompression
@btime jpeg_decode(filename); # 224.760 ms (7 allocations: 51.54 MiB)
# only decompress luminance component
@btime jpeg_decode(Gray, filename); # 91.157 ms (6 allocations: 17.18 MiB)
# process only a few pixels
@btime jpeg_decode(filename; scale_ratio=0.25); # 77.254 ms (8 allocations: 3.23 MiB)
# process only a few pixels for luminance component
@btime jpeg_decode(Gray, filename; scale_ratio=0.25); # 63.119 ms (6 allocations: 1.08 MiB)

An exclusive alternative to scale_ratio is preferred_size:

# minimal `scale_ratio` that output size is greater than or equal to (512, 512)
jpeg_decode(filename; preferred_size=(512, 512)) # size: (751, 750)
# maximal `scale_ratio` that output size is less than or equal to (512, 512)
jpeg_decode(filename; preferred_size=(<=, (512, 512))) # size: (376, 375)


The purpose of this project is to replace ImageMagick.jl with ImageIO. Steven G. Johnson first initialized an early draft version JpegTurbo.jl, this package steals the name from him :). Clang.jl is used to generate the low-level ccall wrapper. Yupei Qi, the current maintainer of Clang.jl, has generously help me debugging C-related codes. This package won't work at all without his help. My another prior project Sixel.jl was also under his generous guidance.