lua-vips

This is a Lua binding for the libvips image processing library. libvips is a fast image processing library with low memory needs. lua-vips uses ffi and needs either

• luajit >= 2.0 or
• standard Lua (5.1 up to 5.4) combined with the luaffi-tkl Lua package.

On the x64 architecture lua-vips is continuously tested

• on Linux and MacOS with Lua 5.1, 5.2, 5.3, 5.4, luajit and openresty-luajit
• on Windows using MSYS2 MinGW-w64 with Lua 5.3, 5.4 and luajit

lua-vips should work on arm64 (recently tested on a Pinephone Pro using PostmarketOS) and possibly x86 (currently untested) as well.

The libvips documentation includes a handy searchable table of every operation in libvips. This is a good place to check if it supports some feature you need. Read on to see how to call libvips operations.

Example

Install the libvips shared library, then install this rock with:

luarocks install lua-vips

When used with LuaJIT please first exhibit luaffi-tkl as provided by the VM via:

luarocks config --lua-version=5.1 rocks_provided.luaffi-tkl 2.1-1

Example:

local vips = require "vips"

-- fast thumbnail generator
local image = vips.Image.thumbnail("somefile.jpg", 128)
image:write_to_file("tiny.jpg")

-- make a new image with some text rendered on it
image = vips.Image.text("Hello <i>World!</i>", {dpi = 300})

-- call a method
image = image:invert()

-- use the `..` operator to join images bandwise
image = image .. image .. image

-- add a constant
image = image + 12
-- add a different value to each band
image = image + { 1, 2, 3 }
-- add two images
image = image + image

-- split bands up again
b1, b2, b3 = image:bandsplit()

-- read a pixel from coordinate (10, 20)
r, g, b = image(10, 20)

-- make all pixels less than 128 bright blue
--    :less(128) makes an 8-bit image where each band is 255 (true) if that 
--        value is less than 128, and 0 (false) if it's >= 128 ... you can use
---       images or {1,2,3} constants as well as simple values
--    :bandand() joins all image bands together with bitwise AND, so you get a
--        one-band image which is true where all bands are true
--    condition:ifthenelse(then, else) takes a condition image and uses true or
--        false values to pick pixels from the then or else images ... then and
--        else can be constants or images
image = image:less(128):bandand():ifthenelse({ 0, 0, 255 }, image)

-- go to Yxy colourspace
image = image:colourspace("yxy")

-- pass options to a save operation
image:write_to_file("x.png", { compression = 9 })

How it works

libvips has quite a bit of introspection machinery built in.

When you call something like image:hough_circle{ scale = 4 }, the __index method on the lua-vips image class opens libvips with ffi and searches for an operation called hough_circle. It discovers what arguments the operation takes, checks you supplied the correct arguments, and transforms them into the form that libvips needs. It executes the operator, then pulls out all the results and returns them as a Lua table.

This means that, although lua-vips supports almost 300 operators, the binding itself is small, should be simple to maintain, and should always be up to date.

Getting more help

The libvips website has a handy table of all the libvips operators. Each one links to the main API docs so you can see what you need to pass to it.

A simple way to see the arguments for an operation is to try running it from the command-line. For example:

$ vips embed
embed an image in a larger image
usage:
   embed in out x y width height
where:
   in           - Input image, input VipsImage
   out          - Output image, output VipsImage
   x            - Left edge of input in output, input gint
			default: 0
			min: -1000000000, max: 1000000000
   y            - Top edge of input in output, input gint
			default: 0
			min: -1000000000, max: 1000000000
   width        - Image width in pixels, input gint
			default: 1
			min: 1, max: 1000000000
   height       - Image height in pixels, input gint
			default: 1
			min: 1, max: 1000000000
optional arguments:
   extend       - How to generate the extra pixels, input VipsExtend
			default: black
			allowed: black, copy, repeat, mirror, white, background
   background   - Color for background pixels, input VipsArrayDouble
operation flags: sequential 

So you can call embed like this:

local image = image:embed(100, 100, image:width() + 200, image:height() + 200,
    { extend = "mirror" })

To add a 100 pixel mirror edge around an image.

Features

This section runs through the main features of the binding.

To load the binding use:

local vips = require "vips"

Make images

You can make images from files or from buffers (Lua strings), you can wrap a vips image around an ffi array, or you can use one of the libvips create operators to make an image for you.

image = vips.Image.new_from_file(filename [, options])

Opens the file and returns an image. You can pass a set of options in a final table argument, for example:

local image = vips.Image.new_from_file("somefile.jpg", 
   { access = "sequential" })

Some options are specific to some file types, for example, shrink, meaning shrink by an integer factor during load, only applies to images loaded via libjpeg.

You can embed options in filenames using the standard libvips syntax. For example, these are equivalent:

local image = vips.Image.new_from_file("somefile.jpg", { shrink = 2 })
local image = vips.Image.new_from_file("somefile.jpg[shrink=2]")

You can call specific file format loaders directly, for example:

local image = vips.Image.jpegload("somefile.jpg", { shrink = 4 })

The loader section in the API docs lists all loaders and their options.

image = vips.Image.new_from_buffer(string [, string_options, options])

The string argument should contain an image file in some container format, such as JPEG. You can supply options, just as with new_from_file. These are equivalent:

local image = vips.Image.new_from_buffer(string, "", { shrink = 2 })
local image = vips.Image.new_from_buffer(string, "shrink=2")

Use (for example) vips.Image.jpegload_buffer to call a loader directly.

image = vips.Image.new_from_memory(data, width, height, bands, format)

This wraps a libvips image around a FFI memory array. The memory array should be formatted as a C-style array. Images are always band-interleaved, so an RGB image three pixels across and two pixels down, for example, is laid out as:

RGBRGBRGB
RGBRGBRGB

Example:

local width = 64
local height = 32
local data = ffi.new("unsigned char[?]", width * height)
local im = vips.Image.new_from_memory(data, width, height, 1, "uchar")

The returned image is using a pointer to the data area, but the Lua/LuaJIT interpreter won't always know this. You should keep a reference to data alive for as long as you are using any downstream images, or you'll get a crash.

image = vips.Image.new_from_memory_ptr(data, size, width, height, bands, format)

Same as new_from_memory, but for any kind of data pointer (non-FFI allocated) by specifying the length of the data in bytes. The pointed data must be valid for the lifespan of the image and any downstream images.

A string can be used as the data pointer thanks to FFI semantics.

image = vips.Image.new_from_image(image, pixel)

Makes a new image with the size, format, and resolution of image, but with each pixel having the value pixel. For example:

local new_image = vips.Image.new_from_image(image, 12)

Will make a new image with one band where every pixel has the value 12. You can call it as a member function. pixel can be a table to make a many-band image, for example:

local new_image = image:new_from_image{ 1, 2, 3 }

Will make a new three-band image, where all the red pixels have the value 1, greens are 2 and blues are 3.

image = vips.Image.new_from_array(array [, scale [, offset]])

Makes a new image from a Lua table. For example:

local image = vips.Image.new_from_array{ 1, 2, 3 }

Makes a one-band image, three pixels across and one high. Use nested tables for 2D images. You can set a scale and offset with two extra number parameters -- handy for integer convolution masks.

local mask = vips.Image.new_from_array(
    {{-1,  -1, -1}, 
     {-1,  16, -1}, 
     {-1,  -1, -1}}, 8)
local image = image:conv(mask, { precision = "integer" })

image = vips.Image.copy_memory(self)

The image is rendered to a large memory buffer, and a new image is returned which represents the memory area.

This is handy for breaking a pipeline.

image = vips.Image.black(width, height)

Makes a new one band, 8 bit, black image. You can call any of the libvips image creation operators in this way, for example:

local noise = vips.Image.perlin(256, 256, { cell_size = 128 })

See:

http://libvips.github.io/libvips/API/current/libvips-create.html

Get and set image metadata

You can read and write aribitrary image metadata.

number = vips.Image.get_typeof(image, field_name)

This returns the GType for a field, or 0 if the field does not exist. vips.gvalue has a set of GTypes you can check against.

mixed = vips.Image.get(image, field_name)

This reads any named piece of metadata from the image, for example:

local version = image:get("exif-ifd2-ExifVersion")

The item is converted to some Lua type in the obvious way. There are convenient shortcuts for many of the standard fields, so these are equivalent:

local width = image:get("width")
local width = image:width()

If the field does not exist, lua-vips will throw an error. Use get_typeof to check for the existence of a field.

vips.Image.set_type(image, gtype, field_name, value)

This creates a new metadata item of the specified type, name and value.

vips.Image.set(image, field_name, value)

This changes the value of an existing field, but will not change its type.

You can't use set() to change core fields such as like width or interpretation. Use copy() instead.

Image references will be shared by the operation cache, so modifying an image can change an image somewhere else in your program. Before changing an image, you must make sure you own a private copy of an image with copy.

local new_image = image:copy()
new_image:set("orientation", 7)

boolean = vips.Image.remove(image, field_name)

This will remove a piece of metadata. It returns true if an item was successfully removed, false otherwise.

As with set, you must use copy before removing a metadata item.

Call any libvips operation

You can call any libvips operation as a member function, for example hough_circle, the circular Hough transform:

http://libvips.github.io/libvips/API/current/libvips-arithmetic.html#vips-hough-circle

Can be called from Lua like this:

local image2 = image:hough_circle{ scale = 2, max_radius = 50 }

The rules are:

1. self is used to set the first required input image argument.

2. If you supply one more argument than the number of required arguments, and the final argument you supply is a table, that extra table is used to set any optional input arguments.

3. If you supply a constant (a number, or a table of numbers) and libvips wants an image, your constant is automatically turned into an image using the first input image you supplied as a guide.

4. For enums, you can supply a number or a string. The string is an enum member nickname (the part after the final underscore).

5. MODIFY arguments, for example the image you pass to draw_circle, are copied to memory before being set, and the new image is returned as one of the results.

6. Operation results are returned as an unpacked array in the order: all required output args, then all optional output args, then all deprecated output args.

You can write (for example):

max_value = image:max()

To get the maximum value from an image. If you look at the max operator, it can actually return a lot more than this. You can write:

max_value, x, y = image:max()

To get the position of the maximum, or:

max_value, x, y, maxes = image:max{ size = 10 }

and maxes will be an array of the top 10 maximum values in order.

Operator overloads

The Lua operators are overloaded in the obvious way, so you can write (for example):

image = (image * 2 + 13) % 4

and the appropriate vips operations will be called. You can mix images, number constants, and array constants freely.

The relational operators are not overloaded, unfortunately; Lua does not permit this. You must write something like:

image = image:less(128):ifthenelse(128, image)

to set all values less than 128 to 128.

__call (ie. ()) is overloaded to call the libvips getpoint operator. You can write:

image = vips.Image.new_from_file("k2.jpg")
r, g, b = image(10, 10)

and r, g, b will be the RGB values for the pixel at coordinate (10, 10).

.. is overloaded to mean bandjoin.

Use im:bands() to get the number of bands and im:extract_band(N) to extract a band (note bands number from zero). lua-vips does not overload # and [] for this, since mixing numbering from zero and one causes confusion.

Convenience functions

A set of convenience functions are also defined.

array<image> = image:bandsplit()

This splits a many-band image into an array of one band images.

image:bandjoin()

The bandjoin operator takes an array of images as input. This can be awkward to call --- you must write:

image = vips.Image.bandjoin{ image, image }

to join an image to itself. Instead, lua-vips defines bandjoin as a member function, so you write:

image = image:bandjoin(image)

to join an image to itself, or perhaps:

image = R:bandjoin{ G, B }

to join three RGB bands. Constants work too, so you can write:

image = image:bandjoin(255)
image = R:bandjoin{ 128, 23 }

The bandrank and composite operators works in the same way.

image = condition_image:ifthenelse(then_image, else_image [, options])

This uses the condition image to pick pixels between then and else. Unlike all other operators, if you use a constant for then_image or else_image, they first match to each other, and only match to the condition image if both then and else are constants.

image = image:sin()

Many vips arithmetic operators are implemented by larger operators which take an enum to set their action. For example, sine is implemented by the math operator, so you must write:

image = image:math("sin")

This is annoying, so a set of convenience functions are defined to enable you to write:

image = image:sin()

There are about 40 of these.

Write

You can write images to files, to ffi arrays, or to formatted strings.

image:write_to_file(filename [, options])

The filename suffix is used to pick the save operator. Just as with new_from_file, not all options will be correct for all file types. You can call savers directly if you wish, for example:

image:jpegsave("x.jpg", { Q = 90 })

string = image:write_to_buffer(suffix [, options])

The suffix is used to pick the saver that is used to generate the result, so ".jpg" will make a JPEG-formatted string. Again, you can call the savers directly if you wish, perhaps:

local str = image:jpegsave_buffer{ Q = 90 }

memory = image:write_to_memory()

A large ffi char array is allocated and the image is rendered to it.

local mem = image:write_to_memory()
print("written ", ffi.sizeof(mem), "bytes to", mem)

ptr, size = image:write_to_memory_ptr()

An allocated char array pointer (GCd with a ffi.gc callback) and the length in bytes of the image data is directly returned from libvips (no intermediate FFI allocation).

Error handling

Most lua-vips methods will call error() if they detect an error. Use pcall() to call a method and catch an error.

Use get_typeof to test for a field of a certain name without throwing an error.

The libvips operation cache

libvips keeps a cache of recent operations, such as load, save, shrink, and so on. If you repeat an operation, you'll get the cached result back.

It keeps track of the number of open files, allocated memory and cached operations, and will trim the cache if more than 100 files are open at once, more than 100mb of memory has been allocated, or more than 1,000 operations are being held.

Normally this cache is useful and harmless, but for some applications you may want to change these values.

-- set number of cached operations
vips.cache_set_max(100)
-- set maximum cache memory use
vips.cache_set_max_mem(10 * 1024 * 1024)
-- set maximum number of open files
vips.cache_set_max_files(10)

Development

Setup for Ubuntu

Configure luarocks for a local tree

luarocks help path

append

eval `luarocks path`
export PATH="$HOME/.luarocks/bin:$PATH"

to ~/.bashrc.

Install

luarocks --local make

Unit testing

You need:

luarocks --local install busted 
luarocks --local install luacov
luarocks --local install say

Then to run the test suite:

busted . 

for verbose output:

busted . -o gtest -v

Linting and static analysis

You need:

luarocks --local install luacheck

Then to run the linter:

luacheck .

Test

Run the example script with:

lua example/hello-world.lua

Update rock

rm *.src.rock
luarocks upload lua-vips-1.1-11.rockspec --api-key=xxxxxxxxxxxxxx

Links

http://luajit.org/ext_ffi_api.html

http://luajit.org/ext_ffi_semantics.html

https://github.com/luarocks/luarocks/wiki/creating-a-rock

https://olivinelabs.com/busted/

Running on Windows using Mingw-w64

Installing lua-vips on Windows is a bit harder than on Unix systems. We recommend using MinGW (Minimalist GNU for Windows) for the installation. Here are the steps:

1. Install MSYS2 to the default path.

2. Start Mingw-w64 64bit console from the start menu. Check that is says MINGW64. The following steps happen in that console.

3. Update MSYS2 using

      pacman -Syuu

4. Install the build tools (including Lua 5.4 and Luarocks) via

      pacman -S git make mingw-w64-x86_64-toolchain mingw-w64-x86_64-lua-luarocks

5. Install libvips with (optional) dependencies via

      pacman -S
         mingw-w64-x86_64-libvips
         mingw-w64-x86_64-openslide
         mingw-w64-x86_64-libheif
         mingw-w64-x86_64-libjxl
         mingw-w64-x86_64-imagemagick
         mingw-w64-x86_64-poppler

6. Optionally: If you want to use lua-vips with LuaJIT instead of Lua 5.4 install LuaJIT via

      pacman -S mingw-w64-x86_64-luajit
      luarocks config --scope system lua_version 5.1
      luarocks config --scope system lua_interpreter luajit.exe
      luarocks config --scope system variables.LUA_DIR /mingw64/bin
      luarocks config --scope system variables.LUA_INCDIR /mingw64/include/luajit-2.1/
      luarocks config --scope system rocks_provided.luaffi-tkl 2.1-1

7. Install lua-vips via

      luarocks install lua-vips

   or clone the repository and run luarocks make in the lua-vips folder.

8. Add C:\msys64\mingw64\bin and C:\msys64\usr\bin to the top of your PATH environment variable in the Windows Advanced system settings and restart the console.

9. Run lua or luajit and try

   vips = require "vips"
   print(vips.Image.xyz(3,2))

Running under Wine (Windows emulation on Linux)

@jcupitt used the luapower all-in-one to get a 64-bit Windows LuaJIT build:

https://luapower.com/

LuaJIT on Windows searches PATH to find DLLs. You can't set this directly from Linux, you have to change the registry. See:

https://www.winehq.org/docs/wineusr-guide/environment-variables

Then add the bin area of the libvips Windows build to PATH.

z:\home\john\GIT\build-win64\8.5\vips-dev-8.5\bin

You must have no trailing backslash.

Try LuaJIT:

$ ~/packages/luajit/luapower-all-master/bin/mingw64/luajit.exe 
LuaJIT 2.1.0-beta2 -- Copyright (C) 2005-2016 Mike Pall.
http://luajit.org/
JIT: ON SSE2 SSE3 SSE4.1 fold cse dce fwd dse narrow loop abc sink fuse
> print(os.getenv("PATH"))
C:\windows\system32;C:\windows;C:\windows\system32\wbem;z:\home\john\GIT\build-win64\8.5\vips-dev-8.5\bin
> ffi = require "ffi"
> ffi.load("libvips-42.dll")
> ^D

The Windows luajit will pick up your .luarocks/share/lua/5.1/vips.lua install, so to test just install and run:

$ ~/packages/luajit/luapower-all-master/bin/mingw64/luajit.exe
LuaJIT 2.1.0-beta2 -- Copyright (C) 2005-2016 Mike Pall. http://luajit.org/
  JIT: ON SSE2 SSE3 SSE4.1 fold cse dce fwd dse narrow loop abc sink fuse
> vips = require "vips"
> x = vips.Image.new_from_file("z:\\data\\john\\pics\\k2.jpg")
> print(x:width())
1450
> x = vips.Image.text("hello", {dpi = 300})
> x:write_to_file("x.png")
>