NEWS.md

v0.7

Add feature_transform and distance_transform

v0.6

Images has been rewritten essentially from scratch for this release. The major goals of the release are:

• More consistent treatment of spatial orientation
• Preserve key properties upon indexing
• Intuitive handling of indexed (colormap) images
• Better support for a wider range of array types
• Improvements in type-stability of many operations
• Improvements in the user experience through easier interfaces, more informative error messages, and improved printing/display
• Improvements in documentation
• For users of former releases of Images, as smooth an upgrade path as can be practically provided, through deprecations or informative error messages

Key changes (of which many are breaking):

• Many properties that were formerly in the dictionary (colorspace, spatial orientation, pixel spacing, and the presence/absence of a time dimension) are now encoded by the type system. The ImageAxes package (a small extension of AxisArrays) is now used for several of these properties. This fixes the former loss of information like spatial orientation when images were indexed (img[5:20, 5:20]).

• The Image type (an array + dictionary) has been renamed ImageMeta, and should be much less needed now that most properties can be encoded with AxisArrays. ImageMeta is still useful if you need to encode information like date/time at which the image was taken, sky coordinates, patient IDs, or experimental conditions. Otherwise, it's recommended to use regular Arrays, or AxisArrays if you need to endow axes with "meaning."

• Full commitment to the use of Colorant types (as defined by the ColorTypes and Colors packages) for encoding color images. Arrays are no longer allowed to declare that they use one axis (dimension) to store color information, i.e., a m×n×3 Float32 array would be displayed as a 3d grayscale image, not an RGB image. This choice facilitates efficient and type-stable indexing behavior and enhances consistency. "Lazy" interconversion between arbitrary numeric arrays and color arrays are provided by two new view types, colorview and channelview, defined in the ImageCore package. These types hopefully remove any awkwardness from the new requirement.

• For an indexed (colormap) image imgi, indexing with imgi[i,j] used to return the index, not the pixel value. This operation now returns the pixel value, with the consequence that indexed images largely act like the array they represent. Indexed images are defined and handled by the IndirectArrays package.

• Image filtering has been greatly improved. imfilter_fft and imfilter_gaussian have both been rolled into imfilter. FFT/FIR filtering is chosen automatically (though a choice can be specified) depending on kernel size, aiming for the best performance in all cases and a consistent interface for specifying defaults. The main filtering algorithms have been considerably improved, particularly for separable kernels, and feature cache-efficient tiling and multithreading. The performance enhancement is as large as 10-fold in some cases, particularly when starting Julia with multiple threads. Certain constructs for specifying boundary conditions have been deprecated and replaced with dispatch-leveraging alternatives. Specification of standard kernels has been changed considerably, and has been split out into two modules, Kernel and KernelFactors, both defined in the ImageFiltering package. In particular note the IIRGaussian types which contain the functionality that was formerly in imfilter_gaussian.

• Nonlinear filtering operations have been added with mapwindow. Among the supported functions is median!, thus providing an implementation of median-filtering.

• Previous versions of Images used reinterpret for several operations, but reinterpret fails for most AbstractArrays other than Array. This release implements alternative mechanisms (e.g., based on the MappedArrays package) that work for any AbstractArray type. (When it would help performance, reinterpret is still used when applicable.) Consequently, this release features better support for a wider range of array types.

• Several improvements have been made to the handling of fixed-point numbers, which permit the use of 8- and 16-bit types that act similarly to floating-point numbers and which permit a consistent criterion for "black" (0.0) and "white" (1.0) independent of storage type. Specifically:

  • Trying to convert out-of-bounds values now gives an informative error message rather than just InexactError
  • Several bugs in FixedPointNumber operations have been fixed, and such operations are more consistent about return types
  • FixedPointNumbers are now printed more compactly

• A new package, ImageTransformations, is underway for rotation, resizing, and other geometric operations on images.

• Many deprecation warnings were designed to help users of the current Images package transition to the new framework.

Other changes (all of which are breaking):

• The gradient components returned by imgradients match the dimensions of the input; in g1, g2, ... = imgradients(img, ...), g1 corresponds to the gradient along the first dimension, g2 along the second, and so on.

• sobel and other filters have been normalized so that the returned "gradient components" are scaled to estimate the actual derivatives. For example, for sobel the normalization factor is 1/8 compared to earlier releases.

• extrema_filter has been deprecated in favor of mapwindow(extrema, A, window). However, this returns an array of (min,max) tuples rather than separate min, max arrays. This is intended to transition towards a future API where one can pass min or max in place of extrema to obtain just one of these. Currently, you can retrieve the min array with first.(mm) and the max array with last.(mm).

• The old extrema_filter discards the edges of the image, whereas the new one (based on mapwindow) returns an array of the same size as the input.

• The output of blob_LoG is now a Vector{BlobLoG}, a new exported immutable, rather than the old tuple format.

• findlocalextrema now returns a Vector{CartesianIndex{N}} rather than a Vector{NTuple{N,Int}}. This makes it ready for use in efficient indexing.

Changes in related packages:

• NRRD.jl has been extensively revamped. The NRRD format lacks an official test suite, and hence it was always uncertain how well the package supported "the standard" (to the extent that there is one). However, it was discovered that unu make can generate files that can serve as a test suite, and using this strategy several incompatibilities in our former version were noted and fixed. It is possible that old .nrrd files written by julia might not be readable without making manual edits to the header.

Older versions

For earlier history, please see the git revision history.