imageTransforms.hpp

Go to the documentation of this file.

/** \file imageTransforms.hpp
 * \author Jared R. Males
 * \brief Image interpolation and transformation
 * \ingroup image_processing_files
 *
 */
 
//***********************************************************************//
// Copyright 2015, 2016, 2017, 2018 Jared R. Males (jaredmales@gmail.com)
//
// This file is part of mxlib.
//
// mxlib is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// mxlib is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with mxlib.  If not, see <http://www.gnu.org/licenses/>.
//***********************************************************************//
 
#ifndef improc_imageTransforms_hpp
#define improc_imageTransforms_hpp
 
#include <cstddef>
#include <cmath>
 
#include <iostream>
#include <limits>
 
namespace mx
{
namespace improc
{
 
/// Transformation by bi-linear interpolation
/** \ingroup image_transforms
 */
template <typename _arithT>
struct bilinearTransform
{
    typedef _arithT arithT;
 
    static const size_t width = 2;
    static const size_t lbuff = 0;
 
    template <typename arrT, typename arithT>
    void operator()( arrT &kern, arithT x, arithT y )
    {
        kern.resize( width, width );
 
        kern( 0, 0 ) = ( 1. - x ) * ( 1. - y );
        kern( 0, 1 ) = ( 1. - x ) * y;
        kern( 1, 0 ) = x * ( 1. - y );
        kern( 1, 1 ) = x * y;
    }
};
 
/// Typedef for bilinearTransform with single precision
/** \ingroup image_transforms
 */
typedef bilinearTransform<float> bilinearTransf;
 
/// Typedef for bilinearTransform with double precision
/** \ingroup image_transforms
 */
typedef bilinearTransform<double> bilinearTransd;
 
/// Transformation by cubic convolution interpolation
/** Uses the cubic convolution interpolation kernel.  See <a
 * href="https://en.wikipedia.org/wiki/Bicubic_interpolation">https://en.wikipedia.org/wiki/Bicubic_interpolation </a>.
 *
 * The parameter \ref cubic should be left as the default -0.5 in most cases, which gives the bicubic spline
 * interpolator.  See <a
 * href="https://en.wikipedia.org/wiki/Cubic_Hermite_spline">https://en.wikipedia.org/wiki/Cubic_Hermite_spline</a>.
 *
 * \tparam _arithT is the type in which to do all calculations.  Should be a floating point type.
 *
 * \test Scenario: Verify direction and accuracy of various image shifts \ref tests_improc_imageTransforms_imageShift
 * "[test doc]"
 *
 * \ingroup image_transforms
 */
template <typename _arithT>
struct cubicConvolTransform
{
    typedef _arithT arithT; ///< The type in which all calculations are performed.
 
    static const size_t width = 4;
    static const size_t lbuff = 1;
 
    arithT cubic{ -0.5 }; ///< The kernel parameter.  The default value -0.5 gives the bicubic spline interpolator.
 
    /// Default c'tor.
    /**
     * This will provide the bicubic spline interpolator
     */
    cubicConvolTransform()
    {
    }
 
    /// Construct setting the kernel parameter.
    explicit cubicConvolTransform( arithT c /**< [in] [optiona] The kernel parameter.  The default value -0.5 gives the bicubic spline interpolator. */)
    {
        cubic = c;
    }
 
    /// Copy c'tor
    cubicConvolTransform( const cubicConvolTransform &t )
    {
        cubic = t.cubic;
    }
 
    /// Calculate the kernel value for a given residual.
    arithT cubicConvolKernel( arithT d )
    {
        if( d <= 1 )
            return ( cubic + 2. ) * d * d * d - ( cubic + 3. ) * d * d + 1.;
 
        if( d < 2 )
            return cubic * d * d * d - 5. * cubic * d * d + 8. * cubic * d - 4. * cubic;
 
        return 0;
    }
 
    ///\todo why is this arithT not just the class's arithT?
    template <typename arrT, typename arithT>
    void operator()( arrT &kern, arithT x, arithT y )
    {
        arithT km2x, km1x, kp1x, kp2x;
        arithT km2y, km1y, kp1y, kp2y;
 
        km2x = cubicConvolKernel( ( 1. + x ) );
        km1x = cubicConvolKernel( x );
        kp1x = cubicConvolKernel( 1. - x );
        kp2x = cubicConvolKernel( 2. - x );
 
        km2y = cubicConvolKernel( ( 1. + y ) );
        km1y = cubicConvolKernel( y );
        kp1y = cubicConvolKernel( 1. - y );
        kp2y = cubicConvolKernel( 2. - y );
 
        kern( 0, 0 ) = km2x * km2y;
        kern( 0, 1 ) = km2x * km1y;
        kern( 0, 2 ) = km2x * kp1y;
        kern( 0, 3 ) = km2x * kp2y;
 
        kern( 1, 0 ) = km1x * km2y;
        kern( 1, 1 ) = km1x * km1y;
        kern( 1, 2 ) = km1x * kp1y;
        kern( 1, 3 ) = km1x * kp2y;
 
        kern( 2, 0 ) = kp1x * km2y;
        kern( 2, 1 ) = kp1x * km1y;
        kern( 2, 2 ) = kp1x * kp1y;
        kern( 2, 3 ) = kp1x * kp2y;
 
        kern( 3, 0 ) = kp2x * km2y;
        kern( 3, 1 ) = kp2x * km1y;
        kern( 3, 2 ) = kp2x * kp1y;
        kern( 3, 3 ) = kp2x * kp2y;
    }
};
 
/// Typedef for cubicConvolTransform with single precision
/** \ingroup image_transforms
 */
typedef cubicConvolTransform<float> cubicConvolTransf;
 
/// Typedef for cubicConvolTransform with double precision
/** \ingroup image_transforms
 */
typedef cubicConvolTransform<double> cubicConvolTransd;
 
/** \ingroup image_transforms
 * @{
 */
 
/// Rotate an image represented as an eigen array
/** Uses the given transformation type to rotate an image.
 *
 * \tparam transformT specifies the transformation to use [will be resolved by compiler]
 * \tparam arrT is the eigen array type of the output [will be resolved by compiler]
 * \tparam arrT2 is the eigen array type of the input [will be resolved by compiler]
 * \tparam floatT is a floating point type [will be resolved by compiler in most cases]
 *
 */
template <typename transformT, typename arrT, typename arrT2, typename floatT>
void imageRotate( arrT &transim,   ///< [out] The rotated image.  Must be pre-allocated.
                  const arrT2 &im, ///< [in] The image to be rotated.
                  floatT dq,       ///< [in] the angle, in radians, by which to rotate in the c.c.w. direction
                  transformT trans ///< [in] is the transformation to use
)
{
    typedef typename transformT::arithT arithT;
    arithT cosq, sinq;
    arithT x0, y0, x, y;
    arithT xcen, ycen;
 
    int Nrows, Ncols;
 
    int i0, j0;
 
    const int lbuff = transformT::lbuff;
    const int width = transformT::width;
 
    cosq = cos( dq );
    sinq = sin( dq );
 
    Nrows = im.rows();
    Ncols = im.cols();
 
    transim.resize( Nrows, Ncols );
 
    // The geometric image center
    xcen = 0.5 * ( Nrows - 1. );
    ycen = 0.5 * ( Ncols - 1. );
 
    int xulim = Nrows - width + lbuff; // - 1;
    int yulim = Ncols - width + lbuff; // - 1;
 
    arithT xc_x_cosq = xcen * cosq;
    arithT xc_x_sinq = xcen * sinq;
    arithT yc_x_cosq = ycen * cosq;
    arithT yc_x_sinq = ycen * sinq;
 
    xc_x_cosq += yc_x_sinq;
    xc_x_sinq -= yc_x_cosq;
 
#ifdef MXLIB_USE_OMP
#pragma omp parallel private( x0, y0, i0, j0, x, y )
#endif
    {
        arithT i_x_cosq, i_x_sinq;
        arrT kern;
        kern.resize( width, width );
 
#ifdef MXLIB_USE_OMP
#pragma omp for schedule( static, 1 )
#endif
        for( int i = 0; i < Nrows; ++i )
        {
            i_x_cosq = i * cosq - xc_x_cosq;      // + xcen;
            i_x_sinq = -( i * sinq - xc_x_sinq ); // + ycen;
 
            for( int j = 0; j < Ncols; ++j )
            {
                // We are actually doing this rotation matrix:
                // x0 =  (i-xcen)*cosq + (j-ycen)*sinq;
                // y0 = -(i-xcen)*sinq + (j-ycen)*cosq;
                // This is the minimum-op representation of the above rotation matrix:
                x0 = i_x_cosq + j * sinq;
                y0 = i_x_sinq + j * cosq;
 
                // Get lower left index
                i0 = x0 + xcen;
                j0 = y0 + ycen;
 
                if( i0 <= lbuff || i0 >= xulim || j0 <= lbuff || j0 >= yulim )
                {
                    transim( i, j ) = 0;
                    continue;
                }
 
                // Get the residual
                x = x0 + xcen - i0;
                y = y0 + ycen - j0;
 
                trans( kern, x, y );
                transim( i, j ) = ( im.block( i0 - lbuff, j0 - lbuff, width, width ) * kern ).sum();
            } // for j
        } // for i
    } // #pragma omp parallel
 
} // void imageRotate(arrT & transim, const arrT2  &im, floatT dq, transformT trans)
 
/// Shift an image by whole pixels with (optional) wrapping.
/** The output image can be smaller than the input image, in which case the wrapping (if enabled) still occurs for the
 * input image, but only output images worth of pixels are actually shifted.  This is useful, for instance, when
 * propagating large turbulence phase screens where one only needs a small section at a time.
 *
 * \tparam outputArrT is the eigen array type of the output [will be resolved by compiler]
 * \tparam inputArrT is the eigen array type of the input [will be resolved by compiler]
 *
 * \test Scenario: Verify direction and accuracy of various image shifts \ref tests_improc_imageTransforms_imageShift
 * "[test doc]"
 */
template <typename outputArrT, typename inputArrT>
void imageShiftWP( outputArrT &out, ///< [out] contains the shifted image.  Must be pre-allocated, but can be smaller
                                    ///< than the in array.
                   inputArrT &in,   ///< [in] the image to be shifted.
                   int dx,          ///< [in] the amount to shift in the x direction
                   int dy,          ///< [in] the amount to shift in the y direction
                   bool wrap = true ///< [in] flag controlling whether or not to wrap around
)
{
    dx %= in.rows();
    dy %= in.cols();
 
    int outr = out.rows();
    int outc = out.cols();
    int inr = in.rows();
    int inc = in.cols();
 
#ifdef MXLIB_USE_OMP
#pragma omp parallel
#endif
    {
        int x, y;
 
        if( wrap )
        {
#ifdef MXLIB_USE_OMP
#pragma omp for
#endif
            for( int cc = 0; cc < outc; ++cc )
            {
                y = cc - dy;
 
                if( y < 0 )
                    y += inc;
                else if( y >= inc )
                    y -= inc;
 
                for( int rr = 0; rr < outr; ++rr )
                {
                    x = rr - dx;
 
                    if( x < 0 )
                        x += inr;
                    else if( x >= inr )
                        x -= inr;
 
                    out( rr, cc ) = in( x, y );
                }
            }
        }
        else
        {
#ifdef MXLIB_USE_OMP
#pragma omp for
#endif
            for( int cc = 0; cc < outc; ++cc )
            {
                y = cc - dy;
 
                if( y < 0 || y >= inc )
                {
                    for( int rr = 0; rr < outr; ++rr )
                    {
                        out( rr, cc ) = 0;
                    }
 
                    continue;
                }
 
                for( int rr = 0; rr < outr; ++rr )
                {
                    x = rr - dx;
 
                    if( x < 0 || x >= inr )
                    {
                        out( rr, cc ) = 0;
                        continue;
                    }
 
                    out( rr, cc ) = in( x, y );
                }
            }
        } // if(wrap)-else
    }
}
 
/// Shift an image by whole pixels, wrapping around, with a scaling image applied to the shifted image.
/** The output image can be smaller than the input image, in which case the wrapping still occurs for the input image,
 * but only output images worth of pixels are actually shifted.  This is useful, for instance, when propagating large
 * turbulence phase screens where one only needs a small section at a time.
 *
 * The scaling is applied to the output image.  The scale image must be the same size as the output image.
 *
 * \tparam outputArrT is the eigen array type of the output [will be resolved by compiler]
 * \tparam inputArrT is the eigen array type of the input [will be resolved by compiler]
 * \tparam scaleArrT is the eigen array type of the scale image [will be resolved by compiler]
 *
 */
template <typename outputArrT, typename inputArrT, typename scaleArrT>
void imageShiftWPScale(
    outputArrT
        &out,      ///< [out] contains the shifted image.  Must be pre-allocated, but can be smaller than the in array.
    inputArrT &in, ///< [in] the image to be shifted.
    scaleArrT &scale, ///< [in] image of scale values applied per-pixel to the output (shifted) image, same size as out
    int dx,           ///< [in] the amount to shift in the x direction
    int dy            ///< [in] the amount to shift in the y direction
)
{
    dx %= in.rows();
    dy %= in.cols();
 
    int outr = out.rows();
    int outc = out.cols();
    int inr = in.rows();
    int inc = in.cols();
#ifdef MXLIB_USE_OMP
// #pragma omp parallel
#endif
    {
        int x, y;
 
#ifdef MXLIB_USE_OMP
// #pragma omp for
#endif
        for( int cc = 0; cc < outc; ++cc )
        {
            y = cc - dy;
 
            if( y < 0 )
                y += inc;
            else if( y >= inc )
                y -= inc;
 
            for( int rr = 0; rr < outr; ++rr )
            {
                x = rr - dx;
 
                if( x < 0 )
                    x += inr;
                else if( x >= inr )
                    x -= inr;
 
                out( rr, cc ) = in( x, y ) * scale( rr, cc );
            }
        }
    }
}
 
/// Shift an image.
/** Uses the given transformation type to shift an image such that objects move by (\p dx,\p dy) pixels.
 * The shift is such that an object located at the coordinate
 * (\p -dx, \p -dy) from the center of the image will be moved to the center of the image.  So to move an object
 * located 2 pixels right (dx) and 2 pixels up (dy) from the center to be at the center, use \p dx = -2, \p dy = -2.
 *
 * Note that this does not treat the edges
 * of the image, determined by the buffer width (lbuff) of the kernel and the size of shift.  If you wish to
 * treat the edges, you must pad the image by at least lbuff+abs(shift) pixels in each direction, and
 * implement a strategy (zeros, mirror, wrap) prior to calling this function.
 *
 * \tparam arrOutT is the Eigen-like array type of the output [will be resolved by compiler]
 * \tparam arrInT is the Eigen-like array type of the input [will be resolved by compiler]
 * \tparam floatT1 is a floating point type [will be resolved by compiler]
 * \tparam floatT2 is a floating point type [will be resolved by compiler]
 * \tparam transformT specifies the transformation to use [will be resolved by compiler]
 *
 * \test Scenario: Verify direction and accuracy of various image shifts \ref tests_improc_imageTransforms_imageShift
 * "[test doc]"
 */
template <typename arrOutT, typename arrInT, typename floatT1, typename floatT2, typename transformT>
void imageShift( arrOutT &transim, ///< [out] Will contain the shifted image.  Will be allocated.
                 const arrInT &im, ///< [in] the image to be shifted.
                 floatT1 dx,       ///< [in] the amount to shift in the x direction
                 floatT2 dy,       ///< [in] the amount to shift in the y direction
                 transformT trans  ///< [in] trans is the transformation to use
)
{
    typedef typename transformT::arithT arithT;
 
    int Nrows, Ncols;
 
    const int lbuff = transformT::lbuff;
    const int width = transformT::width;
 
    // If this is a whole pixel, just do that.
    if( dx == floor( dx ) && dy == floor( dy ) )
        return imageShiftWP( transim, im, dx, dy, false );
 
    Nrows = im.rows();
    Ncols = im.cols();
 
    int xulim = Nrows - width + lbuff;
    int yulim = Ncols - width + lbuff;
 
    transim.resize( Nrows, Ncols );
 
#ifdef MXLIB_USE_OMP
#pragma omp parallel
#endif
    {
        int i0, j0;
        // (rx, ry) is fractional residual of shift
        arithT rx = 1 - ( dx - floor( dx ) );
        arithT ry = 1 - ( dy - floor( dy ) );
 
        arrOutT kern;
        kern.resize( width, width );
        trans( kern, rx, ry );
 
#ifdef MXLIB_USE_OMP
#pragma omp for
#endif
        for( int i = 0; i < Nrows; ++i )
        {
            // (i,j) is position in new image
            // (i0,j0) is integer position in old image
 
            i0 = i - dx;
 
            if( i0 <= lbuff || i0 >= xulim )
            {
                for( int j = 0; j < Ncols; ++j )
                {
                    transim( i, j ) = 0;
                }
                continue;
            }
 
            for( int j = 0; j < Ncols; ++j )
            {
                j0 = j - dy;
 
                if( j0 <= lbuff || j0 >= yulim )
                {
                    transim( i, j ) = 0;
                    continue;
                }
 
                transim( i, j ) = ( im.block( i0 - lbuff, j0 - lbuff, width, width ) * kern ).sum();
            } // for j
        } // for i
    } // #pragam omp
 
} // imageShift
 
/// Magnify an image.
/** Uses the given transformation type to magnify the input image to the size of the output image.
  *
  * Here we assume that the image center is the mxlib standard:
  * \code
    x_center = 0.5*(im.rows()-1);
    y_center = 0.5*(im.cols()-1);
  * \endcode
  * Some care is necessary to prevent magnification from shifting the image with respect to this center.  The main
  result is that the
  * magnification factors (which can be different in x and y) are defined thus:
  * \code
    x_mag = (transim.rows()-1.0) / (im.rows()-1.0);
    y_mag = (transim.cols()-1.0) / (im.cols()-1.0);
  * \endcode
  *
  * Example:
  * \code
    im1.resize(512,512);
    //add image to im1
    im2.resize(1024,1024);
    imageMagnify(im2,im1, cubicConvolTransform<double>());
    \endcode
  * In this exmple, the image in im1 will be magnified by `1023.0/511.0 = 2.002x` and placed in im2.
  *
  * This transform function does not handle edges.  If treatment of edges is desired, you must pad the input
  * image using the desired strategy before calling this function.  Note that the padded-size of the input image
  * will affect the magnification factor.
  *
  * \tparam arrOutT is the eigen array type of the output.
  * \tparam arrInT is the eigen array type of the input.
  * \tparam transformT specifies the transformation to use.
  */
template <typename arrOutT, typename arrInT, typename transformT>
void imageMagnify( arrOutT &transim, ///< [out] contains the magnified image.  Must be pre-allocated.
                   const arrInT &im, ///< [in] is the image to be magnified.
                   transformT trans  ///< [in] is the transformation to use
)
{
    typedef typename transformT::arithT arithT;
 
    arithT x0, y0, x, y;
 
    int Nrows, Ncols;
 
    int i0, j0;
 
    const int lbuff = transformT::lbuff;
    const int width = transformT::width;
 
    Nrows = transim.rows();
    Ncols = transim.cols();
 
    int xulim = im.rows() - lbuff - 1;
    int yulim = im.cols() - lbuff - 1;
 
    arithT x_scale = ( (arithT)im.rows() - 1.0 ) / ( transim.rows() - 1.0 ); // this is 1/x_mag
    arithT y_scale = ( (arithT)im.cols() - 1.0 ) / ( transim.cols() - 1.0 ); // this is 1/y_mag
 
    arithT xcen = 0.5 * ( (arithT)transim.rows() - 1.0 );
    arithT ycen = 0.5 * ( (arithT)transim.cols() - 1.0 );
 
    arithT xcen0 = 0.5 * ( (arithT)im.rows() - 1.0 );
    arithT ycen0 = 0.5 * ( (arithT)im.cols() - 1.0 );
 
    //   #pragma omp parallel private(x0,y0,i0,j0,x,y) num_threads(4)
    {
        arrOutT kern;
        kern.resize( width, width );
 
        for( int j = 0; j < Ncols; ++j )
        {
            // (i,j) is position in new image
            // (x0,y0) is true position in old image
            // (i0,j0) is integer position in old image
            // (x, y) is fractional residual of (x0-i0, y0-j0)
 
            y0 = ycen0 + ( j - ycen ) * y_scale;
            j0 = y0;
 
            if( j0 < lbuff || j0 >= yulim )
            {
                for( int i = 0; i < Nrows; ++i )
                {
                    transim( i, j ) = 0;
                }
                continue;
            }
 
            //       #pragma omp for
            for( int i = 0; i < Nrows; ++i )
            {
                x0 = xcen0 + ( i - xcen ) * x_scale;
                i0 = x0; // just converting to int
 
                if( i0 < lbuff || i0 >= xulim )
                {
                    transim( i, j ) = 0;
                    continue;
                }
 
                // Get the residual
                x = x0 - i0;
                y = y0 - j0;
 
                trans( kern, x, y );
                transim( i, j ) = ( im.block( i0 - lbuff, j0 - lbuff, width, width ) * kern ).sum();
            } // for j
        } // for i
    } // #pragma omp
}
 
/// Magnify an image with the cubic convolution interpolator.
/** Uses the cubic convolution interpolator to magnify the input image to the size of the output image.
 *
 * This is a wrapper for imageMagnify with the transform type specified.
 *
 * \tparam arrOutT is the eigen array type of the output.
 * \tparam arrInT is the eigen array type of the input.
 *
 * \overload
 */
template <typename arrOutT, typename arrInT>
void imageMagnify( arrOutT &transim, ///< [out] contains the magnified image.  Must be pre-allocated.
                   const arrInT &im  ///< [in] is the image to be magnified.
)
{
    return imageMagnify( transim, im, cubicConvolTransform<typename arrInT::Scalar>() );
}
 
/// Re-bin an image using the sum, reducing its size while conserving the total flux.
/** Optionally this can be the mean instead of the sum filter, in which case total flux is not conserved.
 */
template <typename imageOutT, typename imageInT>
int imageRebinSum( imageOutT &imout,     ///< [out] the re-binned image.  Must be allocated to size which is an integer
                                         ///< factor smaller than imin.
                   const imageInT &imin, ///< [in] the image to rebin
                   bool mean = false     ///< [in] if true the output is the mean rather than the sum.
)
{
    int rebin = imin.rows() / imout.rows();
    if( imin.cols() / imout.cols() != rebin )
        return -1;
 
    int N = 1;
    if( mean )
        N = rebin * rebin;
    for( int i = 0; i < imout.rows(); ++i )
    {
        for( int j = 0; j < imout.cols(); ++j )
        {
            imout( i, j ) = imin.block( i * rebin, j * rebin, rebin, rebin ).sum() / N;
        }
    }
 
    return 0;
}
 
/// Re-bin an image using the sum, reducing its size while conserving the total flux.  Records the value and position of
/// the re-binned max pixel.
/** Optionally this can be the mean instead of the sum filter, in which case total flux is not conserved.
 *
 * \overload
 */
template <typename imageOutT, typename imageInT>
int imageRebinSum( imageOutT &imout, ///< [out] the re-binned image.  Must be allocated to size which is an integer
                                     ///< factor smaller than imin.
                   int &xMax,        ///< [out] the x-locatioin of the max pixel
                   int &yMax,        ///< [out] the y-locatioin of the max pixel
                   typename imageOutT::Scalar &pMax, ///< [out] the value of the max pixel
                   const imageInT &imin,             ///< [in] the image to rebin
                   bool mean = false                 ///< [in] if true the output is the mean rather than the sum.
)
{
    int rebin = imin.rows() / imout.rows();
    if( imin.cols() / imout.cols() != rebin )
        return -1;
 
    int N = 1;
    if( mean )
        N = rebin * rebin;
 
    xMax = 0;
    yMax = 0;
    pMax = std::numeric_limits<typename imageOutT::Scalar>::lowest();
 
    for( int i = 0; i < imout.rows(); ++i )
    {
        for( int j = 0; j < imout.cols(); ++j )
        {
            imout( i, j ) = imin.block( i * rebin, j * rebin, rebin, rebin ).sum() / N;
            if( imout( i, j ) > pMax )
            {
                pMax = imout( i, j );
                xMax = i;
                yMax = j;
            }
        }
    }
 
    return 0;
}
 
/// Re-bin an image using the mean.
/** This is a wrapper for imageRebinSum with `mean=true`.
 */
template <typename imageOutT, typename imageInT>
int imageRebinMean( imageOutT &imout,    ///< [out] the re-binned image.  Must be allocated to size which is an integer
                                         ///< factor smaller than imin.
                    const imageInT &imin ///< [in] the image to rebin
)
{
    return imageRebinSum( imout, imin, true );
}
 
/// Re-bin an image using the mean.  Records the value and position of the re-binned max pixel.
/** This is a wrapper for imageRebinSum with `mean=true`.
 *
 * \overload
 */
template <typename imageOutT, typename imageInT>
int imageRebinMean( imageOutT &imout, ///< [out] the re-binned image.  Must be allocated to size which is an integer
                                      ///< factor smaller than imin.
                    int &xMax,        ///< [out] the x-locatioin of the max pixel
                    int &yMax,        ///< [out] the y-locatioin of the max pixel
                    typename imageOutT::Scalar &pMax, ///< [out] the value of the max pixel
                    const imageInT &imin,             ///< [in] the image to rebin
                    bool mean = false                 ///< [in] if true the output is the mean rather than the sum.
)
{
    return imageRebinSum( imout, xMax, yMax, pMax, imin, true );
}
 
/// Re-bin an image, takes the mean with a min/max rejection.
/** The mean is calculated after rejecting the minimuma and maximum value.
 */
template <typename imageOutT, typename imageInT>
int imageRebinMeanReject( imageOutT &imout,    ///< [out] the re-binned image.  Must be allocated to size which is an
                                               ///< integer factor smaller than imin.
                          const imageInT &imin ///< [in] the image to rebin
)
{
    int rebin = imin.rows() / imout.rows();
    if( imin.cols() / imout.cols() != rebin )
        return -1;
 
    int N = rebin * rebin - 2;
 
    for( int i = 0; i < imout.rows(); ++i )
    {
        for( int j = 0; j < imout.cols(); ++j )
        {
            typename imageOutT::Scalar sum = 0;
            typename imageOutT::Scalar max = imin( i * rebin, j * rebin );
            typename imageOutT::Scalar min = imin( i * rebin, j * rebin );
            for( int k = 0; k < rebin; ++k )
            {
                for( int l = 0; l < rebin; ++l )
                {
                    sum += imin( i * rebin + k, j * rebin + l );
                    if( imin( i * rebin + k, j * rebin + l ) > max )
                        max = imin( i * rebin + k, j * rebin + l );
                    if( imin( i * rebin + k, j * rebin + l ) < min )
                        min = imin( i * rebin + k, j * rebin + l );
                }
            }
            imout( i, j ) = ( sum - max - min ) / N; /**/
        }
    }
 
    return 0;
}
 
/// Re-bin an image, takes the mean with a min/max rejection.  Records the value and position of the re-binned max
/// pixel.
/** The mean is calculated after rejecting the minimuma and maximum value.
 *
 * \overload
 */
template <typename imageOutT, typename imageInT>
int imageRebinMeanReject( imageOutT &imout, ///< [out] the re-binned image.  Must be allocated to size which is an
                                            ///< integer factor smaller than imin.
                          int &xMax,        ///< [out] the x-locatioin of the max pixel
                          int &yMax,        ///< [out] the y-locatioin of the max pixel
                          typename imageOutT::Scalar &pMax, ///< [out] the value of the max pixel
                          const imageInT &imin              ///< [in] the image to rebin
)
{
    int rebin = imin.rows() / imout.rows();
    if( imin.cols() / imout.cols() != rebin )
        return -1;
 
    int N = rebin * rebin - 2;
 
    xMax = 0;
    yMax = 0;
    pMax = std::numeric_limits<typename imageOutT::Scalar>::lowest();
 
    for( int i = 0; i < imout.rows(); ++i )
    {
        for( int j = 0; j < imout.cols(); ++j )
        {
            typename imageOutT::Scalar sum = 0;
            typename imageOutT::Scalar max = imin( i * rebin, j * rebin );
            typename imageOutT::Scalar min = imin( i * rebin, j * rebin );
            for( int k = 0; k < rebin; ++k )
            {
                for( int l = 0; l < rebin; ++l )
                {
                    sum += imin( i * rebin + k, j * rebin + l );
                    if( imin( i * rebin + k, j * rebin + l ) > max )
                        max = imin( i * rebin + k, j * rebin + l );
                    if( imin( i * rebin + k, j * rebin + l ) < min )
                        min = imin( i * rebin + k, j * rebin + l );
                }
            }
            imout( i, j ) = ( sum - max - min ) / N;
 
            if( imout( i, j ) > pMax )
            {
                pMax = imout( i, j );
                xMax = i;
                yMax = j;
            }
        }
    }
 
    return 0;
}
 
/// Down-sample an image, reducing its size while conserving the total flux.
/** If the old size is an integer multiple of the new size, this is just a re-bin.  If not an integer multiple,
 * the image is interpolated after performing the closest re-bin, and then re-normalized to conserve flux.
 *
 * \todo Allow selection of interpolator, providing a default version.
 */
template <typename imageOutT, typename imageInT>
void imageDownSample( imageOutT &imout, const imageInT &imin )
{
    typedef typename imageOutT::Scalar Scalar;
 
    // Record this for normalization later
    Scalar inputTotal = fabs( imin.sum() );
 
    // As a first step, rebin to nearest whole pixel factor which is larger than the desired output size
    int closestRebin = imin.rows() / imout.rows(); //, imin.cols()/imout.cols() );
 
    float sample = ( (float)imin.rows() ) / closestRebin;
 
    while( sample != floor( sample ) )
    {
        --closestRebin;
        if( closestRebin == 1 )
            break;
        sample = ( (float)imin.rows() ) / closestRebin;
    }
 
    // Eigen::Array<Scalar, Eigen::Dynamic, Eigen::Dynamic> temp;
    imageOutT temp;
    temp.resize( imin.rows() / closestRebin, imin.cols() / closestRebin );
 
    for( int i = 0; i < temp.rows(); ++i )
    {
        for( int j = 0; j < temp.cols(); ++j )
        {
            temp( i, j ) = imin.block( i * closestRebin, j * closestRebin, closestRebin, closestRebin ).sum();
        }
    }
 
    // If the output image is now the requested size return.
    if( temp.rows() == imout.rows() && temp.cols() == imout.cols() )
    {
        imout = temp;
        return;
    }
    // Otherwise, re-sample using bilinear interpolation.
    typedef bilinearTransform<Scalar> transformT;
 
    transformT trans;
    // Eigen::Array<Scalar, -1,-1> kern;
    imageOutT kern;
 
    const int lbuff = transformT::lbuff;
    const int width = transformT::width;
 
    for( int i = 0; i < imout.rows(); ++i )
    {
        for( int j = 0; j < imout.cols(); ++j )
        {
            double x = ( (double)i / imout.rows() ) * temp.rows();
            double y = ( (double)j / imout.cols() ) * temp.cols();
 
            trans( kern, x - floor( x ), y - floor( y ) );
 
            imout( i, j ) = ( temp.block( floor( x ) - lbuff, floor( y ) - lbuff, width, width ) * kern ).sum();
        }
    }
 
    // Normalize
    Scalar outputTotal = fabs( imout.sum() );
    imout *= inputTotal / outputTotal;
}
 
///@}
 
} // namespace improc
} // namespace mx
 
#endif // improc_imageTransforms_hpp