NEWS

mmand 1.6.3 (2023-02-07)

Too-eager updating of data in-place could lead to errors in some elements of a multidimensional distance transform. This has been corrected. (Reported by Fabio Boschetti.)

display() tests are now run from the R temporary directory, to avoid cluttering up the package directory.
Distance transform tests are now skipped on Solaris, due to irreproducible failures on CRAN.

A separable distance transform is now available through the new distanceTransform() function. This maps the distances from each array element to the nearest boundary between foreground and background, and can be useful for estimating the width of image features. (Hat-tip to Chris Rorden and Philip Rideout.)
It is now possible to suppress the creation of a new device by display(), by setting the "mmand.display.newDevice" option to FALSE.
Tests have been migrated from the "testthat" framework to "tinytest", and coverage has been improved.

The range attribute is now properly updated when an image is binarised within the threshold() function. (Reported by @GhostValk, issue #5.)
Compile-time tweaks to meet updated CRAN requirements.

A configure script is now used to check for OpenMP support on non-Windows platforms. This can be helpful in certain build environments on macOS.

Resampling with the box kernel should no longer produce an erroneous "chequered" pattern. (Reported by Steffen Ehrmann, issue #3.)

The connected components algorithm will no longer erroneously wrap around image boundaries. (Reported by Steffen Ehrmann, issue #2.)

The five-lobe Lanczos windowed sinc kernel function is now available through the lanczosKernel() function, and may be used for resampling.
The new sketch() function can be used to show a compact ASCII representation of a 2D greyscale image or other matrix.
Experimental support for image skeletonisation has been added through the skeletonise() function and its alias, skeletonize(). Three different algorithms are available, with different limitations. Please see the README or ?skeletonise for details.
Logical data is now accepted by package functions, in addition to integer and double-valued data.
The display() function now handles 2D objects of class "array", and respects image attributes set by the "loder" package.
The morph() function gains a renormalise option, and some new operators.
Infinite, NaN and NA values are now ignored by threshold() when using the k-means method (since kmeans() itself does not handle them gracefully). Elements with any of these values are replaced with NAs in the result.
The binary() function now returns the unique value in an attribute, if applicable.
The binarise() function may now also be spelled binarize().
A new binary test image, of an upper-case B, has been added.
The "loder" package is now used instead of "png" for reading PNG images.
The package no longer depends on "reportr".

The C++ back-end for resample() has been rewritten to improve speed and fix implementation issues with cubic spline interpolation. This function (only) will now also be parallelised using OpenMP, if your compiler supports it.
Dilation now works properly with asymmetric kernels. (Reported by Marcin Skowronek.)
The package no longer depends on RcppEigen.

The package can now find connected components in images, representing contiguous subregions, using the new components() function. The LEMON network library is used to solve the associated graph problem.
The Sobel-Feldman kernel and filter are now available through new functions. These can be used to obtain a finite-difference gradient.
The new symmetric() function reports whether or not an array is symmetric.
Gaussian smoothing in multiple dimensions has been made much faster by exploiting the separability of the kernel. Sigma values of zero may also be used in some dimensions. (Hat-tip to Tim Tierney.)
Erosion and dilation did not work as documented when the kernel had a zero at its centre. This has been corrected. (Reported by Scott Flanagan.)
Zero-sum kernels are now properly handled.

A function has been added for thresholding arrays, either using a literal threshold value, or implicitly using k-means clustering.
The display() function now supports multichannel images (e.g., RGB, RGBA). Zero values are also now set to NA when adding to an existing image, to make those pixels transparent.

The README.md file has been expanded into a full tutorial for the package, complete with visual examples.
Calls to the morph() function with merge="sum" previously failed to adjust final pixel values when the whole kernel was not used, notably at the edges of images. This led to values lower than expected. The bug has now been corrected.
The test image has been replaced by a less traditional, but also less sexist, alternative.
Package tests have been updated to make use of new functionality in version 0.9 of the "testthat" package.

The back-end code for the package has been completely rewritten, using C++ and Rcpp rather than C.
The package can now perform image resampling, via resample(). This requires a different type of kernel but is conceptually related to the existing functions based on morph().
The morph() function has been generalised to allow for various new operations. In the process, the concepts of "brush" and "eraser" kernels have been removed.
2D arrays (i.e. matrices or 2D images) can now be visualised using the new display() function.
Functions for mean and median filtering have been added.
There are additional functions for establishing whether an array is binary, and for obtaining information about the structure of a neighbourhood within an array.
Kernels now have a plot() method.
The classical image processing test image, "lena", has been added to the package.
A suite of tests has been added.

Brush kernels which include NAs, notably diamond-shaped kernels, are now properly handled by the C code. The value of the output array was sometimes overwritten by that of the input array when it should not have been.

A mistake in internal memory management which could lead to a segmentation fault has been corrected.