Hi, ITK developers,
I am working for implementing a fast convolution process and I found some thing seems incorrect in the itkMaskNeighborhoodOperatorImageFilter . I made a MaskConvolutionImageFilter using itkMaskNeighborhoodOperatorImageFilter, so that convolution process is able to be faster by applying a small mask as a input. However, regarding my testing results, I found the process still requires 8 to 11 sec to finish convolution even I input only a blank mask. After I trace source code of MaskNeighborhoodOperatorImageFilter. The behavior of the source code does really confuse me, because itk::ConstNeighborhoodIterator do check boundary hitting for each dimension when every time iterator moving. This check action seems very expensive and slow down entire convolution process because it is performed even the target pixel isn’t belong to mask. I have appiled some modification to improve performance of convolution process. Both original code and modified code are shown as follows:
Original code:
ConstNeighborhoodIterator< InputImageType > bit;
typename FaceListType::iterator fit;
ImageRegionIterator< OutputImageType > it;
ImageRegionConstIterator< MaskImageType > mit;
for ( fit = faceList.begin(); fit != faceList.end(); ++fit )
{
bit =
ConstNeighborhoodIterator< InputImageType >(noperator.GetRadius(),
input, *fit);
bit.OverrideBoundaryCondition( this->GetBoundaryCondition() );
bit.GoToBegin();
it = ImageRegionIterator< OutputImageType >(output, *fit);
mit = ImageRegionConstIterator< MaskImageType >(mask, *fit);
while ( !bit.IsAtEnd() )
{
if ( mit.Get() )
{
// Compute the inner product at this pixel
it.Value() = static_cast< typename OutputImageType::PixelType >( smartInnerProduct(bit, noperator) );
}
else
{
// Use the default value or the input value
it.Value() = m_UseDefaultValue ? m_DefaultValue : bit.GetCenterPixel();
}
++bit; // This movement cause the performance issue
++it;
++mit;
progress.CompletedPixel();
}
}
The small modification:
itk::ConstNeighborhoodIterator bit;
itk::ImageRegionConstIterator light_bit;
typename FaceListType::iterator fit;
itk::ImageRegionIterator it;
itk::ImageRegionConstIteratorWithIndex mit;
for(fit = faceList.begin(); fit != faceList.end(); ++fit)
{
bit =
itk::ConstNeighborhoodIterator(noperator.GetRadius(),
input, *fit);
bit.OverrideBoundaryCondition(this->GetBoundaryCondition());
bit.GoToBegin();
light_bit = itk::ImageRegionConstIterator(input, *fit);
it = itk::ImageRegionIterator(output, *fit);
mit = itk::ImageRegionConstIteratorWithIndex(mask, *fit);
for(mit = mit.Begin(); mit != mit.End(); ++mit)
{
if (mit.Get())
{
// Compute the inner product at this pixel
bit.SetLocation( mit.GetIndex() ); // Only check boundary hitting and warp offset when pixels belong to mask area
it.Value() = static_cast(smartInnerProduct(bit, noperator));
}
else
{
// Use the default value or the input value
it.Value() = Superclass::GetUseDefaultValue() ? Superclass::GetDefaultValue() : light_bit.Value();
}
++light_bit;
++it;
}
}
