Changes DIPlib 3.1.0

Changes to DIPlib

New functionality

• Added dip::MarginalPercentile() histogram-based statistics function.

• Added dip::SplitRegions(), a function to transform labeled images.

• Added dip::FlushToZero(), a function to remove denormal values from an image.

• Added dip::maximum_gauss_truncation(), returning the maximum truncation that is useful for Gaussian functions.

• dip::NeighborList allows accessing neighbors by index, through the new member functions Coordinates, Distance and IsInImage.

• Added dip::Measurement::SetObjectIDs().

• Added dip::Image::SwapBytesInSample() to convert from little endian to big endian representation.

• Added dip::ImageReadNPY(), dip::ImageReadNPYInfo(), dip::ImageIsNPY() and dip::ImageWriteNPY(), to work with NumPy’s NPY files, expanding interoperability with Python.

• Compound assignment operators defined for image views (img[ 0 ] += ... now works).

• Added dip::GetLabelBoundingBox().

• Added dip::IsotropicDilation(), dip::IsotropicErosion(), dip::IsotropicClosing() and dip::IsotropicOpening() for binary images.

• Added dip::Image::Sample::Maximum() and dip::Image::Sample::Minimum().

• Added dip::ImposeMinima().

• Added dip::Toggle(), another variant to dip::Select().

• Added dip::MorphologicalSharpening(), based on dip::Toggle().

• It is now possible to construct a default-initialized dip::Histogram, which can only be assigned to. The new function dip::Histogram::IsInitialized() returns false for such an object.

Changed functionality

• Replaced the use of std::map and std::set in various functions to tsl::robin_map and tsl::robin_set, which are significantly faster. This affects dip::Measurement and functions using it (including dip::MeasurementTool::Measure()), dip::GetImageChainCodes(), dip::GetObjectLabels(), dip::Relabel(), dip::ChordLength() and dip::PairCorrelation().

• The speed of the following functions has been improved:
  • dip::Measurement::AddObjectIDs() and dip::Measurement::operator+().
  • dip::Label() (slightly more efficient for 3D and higher-dimensional images).
  • dip::MomentAccumulator.
  • dip::Dilation() and the other basic morphological operators with a 2D diamond or octagonal structuring element, or with a small 2D disk.
  • dip::MorphologicalReconstruction() and functions that depend on it.
  • dip::SeededWatershed() and dip::CompactWatershed(), and functions that depend on them.
  • dip::Thickening() and dip::Thinning().

• dip::OptimalFourierTransformSize() has a new option to return a smaller or equal size, rather than a larger or equal size, so we can crop an image for efficient FFT instead of padding.

• All the variants of the Gaussian filter (dip::Gauss() et al.) now limit the truncation value to avoid unnecessarily large filter kernels. dip::DrawBandlimitedPoint(), dip::DrawBandlimitedBall(), dip::DrawBandlimitedBox(), dip::GaussianEdgeClip() and dip::GaussianLineClip() also limit the truncation in the same way.

• dip::ColorSpaceConverter has a SetWhitePoint member function. dip::ColorSpaceManager::SetWhitePoint calls the SetWhitePoint member function for all registered converters. This allows custom converters to use the white point as well.

• dip::ColorSpaceManager::XYZ is deprecated, use dip::XYZ instead.

• dip::SetNumberOfThreads() and dip::GetNumberOfThreads() set a thread-local value now. Each thread should set its own limit for how many threads DIPlib can use.

• All functions that used randomness internally but didn’t have a dip::Random input parameter now do have such a parameter. Overloaded functions with the old signature create a default-initialized dip::Random object and call the function with the new signature. This affects the following functions: dip::ChordLength(), dip::PairCorrelation(), dip::ProbabilisticPairCorrelation(), dip::Semivariogram(), dip::CostesSignificanceTest(), dip::StochasticWatershed(), dip::KMeansClustering(), and dip::Superpixels().

• dip::Image::Rotation90() can now be called without any arguments at all.

Bug fixes

• dip::Image::ResetNonDataProperties() incorrectly set the number of tensor elements to 1.

• dip::Image::Copy() would incorrectly copy the external interface if the destination image didn’t have one defined.

• dip::Image::HasSameDimensionOrder() didn’t properly ignore singleton dimensions.

• It was possible to forge a new image with normal strides for which dip::Image::HasNormalStrides() was false.

• dip::DrawPolygon2D() produced wrong results for filled polygons when vertices were very close together (distances smaller than a pixel).

• dip::ColorSpaceManager didn’t register the ICH and ISH color spaces.

• dip::MaximumPixel() and dip::MinimumPixel() produced an empty output array when the input was all NaN.

• dip::GaussianMixtureModelThreshold() passed wrong parameter values to dip::GaussianMixtureModel().

• dip::ResampleAt(in, map) didn’t use the pixels at the right or bottom edges of the input image.

• Fixed various issues in underlying code for dip::AreaOpening(), dip::AreaClosing(), dip::VolumeOpening() and dip::VolumeClosing(). These functions now behave as described in the documentation.

• There was a strange rounding error when creating disk-shaped filter kernels and structuring elements, for some even integer sizes, which caused these kernels to be not symmetric over 90 degree rotations.

• dip::MeanAbs() and dip::SumAbs() could produce wrong results for complex inputs.

• libics had a typo that caused out-of-bounds read (#81).

• Fixed error handling in dip::ImageReadJPEG() and dip::ImageWriteJPEG(), which previously would crash when libjpeg produced an error (#80).

• Improved error messages in dip::ImageRead().

• dip::Closing() and dip::Opening() for elliptic and custom structuring elements now have a better behavior at the image border.

• Morphological operations with the "fast line" structuring element triggered an assertion for some lines in 3D and higher dimensions.

• Morphological operations with the isotropic "diamond" structuring element in 3D and higher dimensions was implemented using a set of line structuring elements, but their composition didn’t lead to a ball in city-block distance.

• dip::MorphologicalThreshold(), dip::MorphologicalGist(), dip::MorphologicalRange(), dip::MorphologicalSmoothing(), dip::MultiScaleMorphologicalGradient() and dip::MorphologicalLaplace() all produced wrong output for some combinations of input image data types and option settings.

• Avoid the C++ Standard Library generating an assertion error in debug mode when generating a Poisson-distributed random value with a non-positive mean.

Changes to DIPimage

New functionality

• Added dip_image/ftz, a function to remove denormal values from an image.

Changed functionality

• Most functions now also accept a string array instead of char vector or a cell array of char vectors (that is, "foo" is now interpreted the same as 'foo', and ["foo","bar"] is now interpreted the same as {'foo','bar'}).

• mat2im also accepts -1 as second argument, taking the last array dimension as the tensor dimension.

(See also changes to DIPlib.)

Bug fixes

• readim and writeim, when using MATLAB’s functionality, would work in linear RGB space, instead of sRGB like the DIPlib functions do.

• readics tried to read the file as a TIFF file instead of an ICS file.

• Code like img * [1,2,3] converted the second argument into a 0D column vector image. The same code with e.g. a 3x3 matrix would convert the matrix to a scalar 3x3 image, rather than a 0D matrix image. These now behave more similarly to how they behaved in DIPimage 2.9.

• Assignments into multiple tensor dimensions like img{2:3} = [4,5] or img{2:3} = gaussf(img{2:3}) produced an error message.

• im2mat again puts the tensor dimension (and color channels) as the last dimension of the output array, as it did in DIPimage 2.9.

(See also bugfixes to DIPlib.)

Changes to PyDIP

New functionality

• Added dip.FlushToZero(), dip.SplitRegions(), dip.ApplyWindow().

• Added dip.ImageReadNPY(), dip.ImageReadNPYInfo(), dip.ImageIsNPY() and dip.ImageWriteNPY().

• Added dip.FourierMellinMatch2Dparams(), identical to dip.FourierMellinMatch2D() but it also returns the transform parameters (which dip.FourierMellinMatch2D() should have done from the beginning).

• Added dip.GetLabelBoundingBox().

• Added dip.IsotropicDilation(), dip.IsotropicErosion(), dip.IsotropicClosing() and dip.IsotropicOpening() for binary images.

• Added dip.ImposeMinima(), dip.Toggle() and dip.MorphologicalSharpening().

• Added dip.libraryInformation and dip.__version__.

Changed functionality

• dip.ChordLength, dip.PairCorrelation, dip.ProbabilisticPairCorrelation, dip.Semivariogram, dip.CostesSignificanceTest, dip.StochasticWatershed, dip.KMeansClustering, and dip.Superpixels now use the global random number generator instead of creating a default-initialized one.

(See also changes to DIPlib.)

Bug fixes

• Updated default connectivity, end pixel condition, and boundary condition values for some functions in the morphology and binary image processing modules to match the defaults in DIPlib.

• Updated Linux wheels to correctly package FreeGLUT

(See also bugfixes to DIPlib.)

Changes to DIPjavaio

Bug fixes

• Signed integer images could not be read.