Smooth RGB Image Using Curvature Flow#

Synopsis#

Smooth an RGB image using curvature flow.

Results#

../../../../_images/SmoothRGBImageUsingCurvatureFlow.png — Output In VTK Window#

Code#

C++#

#include "itkImageAdaptor.h"
#include "itkImageRegionIterator.h"
#include "itkNthElementImageAdaptor.h"
#include "itkCurvatureFlowImageFilter.h"
#include "itkComposeImageFilter.h"
#include "itkImage.h"
#include "itkImageFileReader.h"
#include "itkRGBPixel.h"

#include <sstream>

#ifdef ENABLE_QUICKVIEW
#  include "QuickView.h"
#endif

int
main(int argc, char * argv[])
{
  // Verify command line arguments
  if (argc < 2)
  {
    std::cerr << "Usage: " << std::endl;
    std::cerr << argv[0] << " InputImageFile [iterations]" << std::endl;
    return EXIT_FAILURE;
  }

  int iterations = 5;
  if (argc > 2)
  {
    std::stringstream ss(argv[2]);
    ss >> iterations;
  }
  std::string inputFileName = argv[1];

  // Setup types
  using ComponentType = float;
  using PixelType = itk::RGBPixel<ComponentType>;
  using RGBImageType = itk::Image<PixelType, 2>;
  using ImageType = itk::Image<ComponentType, 2>;
  using ImageAdaptorType = itk::NthElementImageAdaptor<RGBImageType, unsigned char>;
  using ComposeType = itk::ComposeImageFilter<ImageType, RGBImageType>;
  using CurvatureFlowType = itk::CurvatureFlowImageFilter<ImageAdaptorType, ImageType>;

  const auto input = itk::ReadImage<RGBImageType>(inputFileName);

  // Run the filter for each component
  auto rAdaptor = ImageAdaptorType::New();
  rAdaptor->SelectNthElement(0);
  rAdaptor->SetImage(input);

  auto rCurvatureFilter = CurvatureFlowType::New();
  rCurvatureFilter->SetInput(rAdaptor);
  rCurvatureFilter->SetNumberOfIterations(iterations);
  rCurvatureFilter->Update();

  auto gAdaptor = ImageAdaptorType::New();
  gAdaptor->SelectNthElement(1);
  gAdaptor->SetImage(input);

  auto gCurvatureFilter = CurvatureFlowType::New();
  gCurvatureFilter->SetInput(gAdaptor);
  gCurvatureFilter->SetNumberOfIterations(iterations);
  gCurvatureFilter->Update();

  auto bAdaptor = ImageAdaptorType::New();
  bAdaptor->SelectNthElement(2);
  bAdaptor->SetImage(input);

  auto bCurvatureFilter = CurvatureFlowType::New();
  bCurvatureFilter->SetInput(bAdaptor);
  bCurvatureFilter->SetNumberOfIterations(iterations);
  bCurvatureFilter->Update();

  // compose an RGB image from the three filtered images

  auto compose = ComposeType::New();
  compose->SetInput1(rCurvatureFilter->GetOutput());
  compose->SetInput2(gCurvatureFilter->GetOutput());
  compose->SetInput3(bCurvatureFilter->GetOutput());

#ifdef ENABLE_QUICKVIEW
  QuickView viewer;
  viewer.AddRGBImage(input.GetPointer(), true, itksys::SystemTools::GetFilenameName(inputFileName));

  std::stringstream desc;
  desc << "CurvatureFlow iterations = " << iterations;
  viewer.AddRGBImage(compose->GetOutput(), true, desc.str());

  viewer.Visualize();
#endif
  return EXIT_SUCCESS;
}

Classes demonstrated#

template<typename TInputImage, typename TOutputImage> class CurvatureFlowImageFilter : public itk::DenseFiniteDifferenceImageFilter<TInputImage, TOutputImage>

Denoise an image using curvature driven flow.

CurvatureFlowImageFilter implements a curvature driven image denoising algorithm. Iso-brightness contours in the grayscale input image are viewed as a level set. The level set is then evolved using a curvature-based speed function:

$I_t = \kappa |\nabla I|$

where $\kappa$ is the curvature.

The advantage of this approach is that sharp boundaries are preserved with smoothing occurring only within a region. However, it should be noted that continuous application of this scheme will result in the eventual removal of all information as each contour shrinks to zero and disappear.

Note that unlike level set segmentation algorithms, the image to be denoised is already the level set and can be set directly as the input using the SetInput() method.

This filter has two parameters: the number of update iterations to be performed and the timestep between each update.

The timestep should be “small enough” to ensure numerical stability. Stability is guarantee when the timestep meets the CFL (Courant-Friedrichs-Levy) condition. Broadly speaking, this condition ensures that each contour does not move more than one grid position at each timestep. In the literature, the timestep is typically user specified and have to manually tuned to the application.

This filter make use of the multi-threaded finite difference solver hierarchy. Updates are computed using a CurvatureFlowFunction object. A zero flux Neumann boundary condition when computing derivatives near the data boundary.

This filter may be streamed. To support streaming this filter produces a padded output which takes into account edge effects. The size of the padding is m_NumberOfIterations on each edge. Users of this filter should only make use of the center valid central region.

Reference: “Level Set Methods and Fast Marching Methods”, J.A. Sethian, Cambridge Press, Chapter 16, Second edition, 1999.

Warning: This filter assumes that the input and output types have the same dimensions. This filter also requires that the output image pixels are of a floating point type. This filter works for any dimensional images.

Input/Output Restrictions: TInputImage and TOutputImage must have the same dimension. TOutputImage’s pixel type must be a real number type.

See: DenseFiniteDifferenceImageFilter
See: CurvatureFlowFunction
See: MinMaxCurvatureFlowImageFilter
See: BinaryMinMaxCurvatureFlowImageFilter

ITK Sphinx Examples:

All ITK Sphinx Examples

Smooth Image Using Curvature Flow

Smooth RGB Image Using Curvature Flow

Subclassed by itk::MinMaxCurvatureFlowImageFilter< TInputImage, TOutputImage >

See itk::CurvatureFlowImageFilter for additional documentation.