Using PARFOR to Speed Up an Image Contrast Enhancement Algorithm

Open Live Script

This example shows how to generate a standalone C library from MATLAB® code that applies a simple histogram equalization function to images to improve image contrast. The example uses parfor to process each of the standard three RGB image planes on separate threads. The example also shows how to generate and run a MEX function in MATLAB prior to generating C code to verify that the MATLAB code is suitable for code generation.

MATLAB Coder™ uses the OpenMP portable shared memory parallel programming standard to implement its support for parfor. See The OpenMP API Specification for Parallel Programming. Whereas MATLAB supports parfor by creating multiple worker sessions, MATLAB Coder uses OpenMP to create multiple threads running on the same machine.

Prerequisites

In order to support parallelization, the compiler must support the OpenMP shared memory parallel programming standard. If your compiler does not have this support, then you can still run this example, but the generated code will run serially.

About the `histequalize` Function

The histequalize.m function takes an image (represented as an NxMx3 matrix) and returns an image with enhanced contrast.

type histequalize

function equalizedImage = histequalize(originalImage) %#codegen
% equalizedImage = histequalize(originalImage)
% Histogram equalization (or linearization) for improving image contrast.
% Given an NxMx3 image, equalizes the histogram of each of the three image
% planes in order to improve image contrast.

    assert(size(originalImage,1) <= 8192);
    assert(size(originalImage,2) <= 8192);
    assert(size(originalImage,3) == 3);
    assert(isa(originalImage, 'uint8'));

    [L, originalHist] = computeHistogram(originalImage);
    equalizedImage = equalize(L, originalHist, originalImage);
end

function [L, originalHist] = computeHistogram(originalImage)
    L = double(max(max(max(originalImage)))) + 1;
    originalHist = coder.nullcopy(zeros(3,L));
    sz = size(originalImage);
    N = sz(1);
    M = sz(2);
    parfor plane = 1:sz(3)
        planeHist = zeros(1,L);
        for y = 1:N
            for x = 1:M
                r = originalImage(y,x,plane);
                planeHist(r+1) = planeHist(r+1) + 1;
            end
        end
        originalHist(plane,:) = planeHist;
    end
end

function equalizedImage = equalize(L, originalHist, originalImage)
    equalizedImage = coder.nullcopy(originalImage);
    sz = size(originalImage);
    N = sz(1);
    M = sz(2);
    normalizer = (L - 1)/(N*M); 
    parfor plane = 1:sz(3)
        planeHist = originalHist(plane,:);
        for y = 1:N
            for x = 1:M               
                r = originalImage(y,x,plane);
                s = 0;
                for j = 0:int32(r)
                    s = s + planeHist(j+1);
                end
                s = normalizer * s;
                equalizedImage(y,x,plane) = s;
            end
        end
    end
end

Generate the MEX Function

Generate a MEX function using the codegen command.

codegen histequalize

Code generation successful.

Before generating C code, you should first test the MEX function in MATLAB to ensure that it is functionally equivalent to the original MATLAB code and that no run-time errors occur. By default, codegen generates a MEX function named histequalize_mex in the current folder. This allows you to test the MATLAB code and MEX function and compare the results.

Read in the Original Image

Use the standard imread command to read the low-contrast image.

lcIm = imread('LowContrast.jpg');
image(lcIm);

Figure contains an axes object. The axes object contains an object of type image.

Run the MEX Function (The Histogram Equalization Algorithm)

Pass the low-contrast image.

hcIm = histequalize_mex(lcIm);

Display the Result

image(hcIm);

Figure contains an axes object. The axes object contains an object of type image.

Generate Standalone C Code

codegen -config:lib histequalize

Code generation successful.

Using codegen with the -config:lib option produces a standalone C library. By default, the code generated for the library is in the folder codegen/lib/histequalize/.

Inspect the Generated Function

Notice that the generated code contains OpenMP pragmas that control parallelization of the code using multiple threads.

type codegen/lib/histequalize/histequalize.c

/*
 * File: histequalize.c
 *
 * MATLAB Coder version            : 24.2
 * C/C++ source code generated on  : 05-Sep-2024 13:47:32
 */

/* Include Files */
#include "histequalize.h"
#include "histequalize_data.h"
#include "histequalize_emxutil.h"
#include "histequalize_initialize.h"
#include "histequalize_types.h"
#include "minOrMax.h"
#include "omp.h"
#include <math.h>
#include <string.h>

/* Function Declarations */
static double computeHistogram(const emxArray_uint8_T *originalImage,
                               double originalHist_data[],
                               int originalHist_size[2]);

static void equalize(double L, const double originalHist_data[],
                     const emxArray_uint8_T *originalImage,
                     emxArray_uint8_T *equalizedImage);

static double rt_roundd_snf(double u);

/* Function Definitions */
/*
 * Arguments    : const emxArray_uint8_T *originalImage
 *                double originalHist_data[]
 *                int originalHist_size[2]
 * Return Type  : double
 */
static double computeHistogram(const emxArray_uint8_T *originalImage,
                               double originalHist_data[],
                               int originalHist_size[2])
{
  double planeHist_data[256];
  double L;
  int tmp_size[3];
  int M;
  int N;
  int i;
  int loop_ub;
  int plane;
  int x;
  int y;
  unsigned char tmp_data[24576];
  unsigned char uv[3];
  const unsigned char *originalImage_data;
  unsigned char L_tmp;
  unsigned char r;
  originalImage_data = originalImage->data;
  maximum(originalImage, tmp_data, tmp_size);
  b_maximum(tmp_data, tmp_size, uv);
  L_tmp = c_maximum(uv);
  L = (double)L_tmp + 1.0;
  originalHist_size[0] = 3;
  originalHist_size[1] = L_tmp + 1;
  N = originalImage->size[0];
  M = originalImage->size[1];
#pragma omp parallel for num_threads(omp_get_max_threads()) private(           \
        r, planeHist_data, loop_ub, y, x, i)

  for (plane = 0; plane < 3; plane++) {
    loop_ub = (int)L;
    memset(&planeHist_data[0], 0, (unsigned int)loop_ub * sizeof(double));
    for (y = 0; y < N; y++) {
      for (x = 0; x < M; x++) {
        r = originalImage_data[(y + originalImage->size[0] * x) +
                               originalImage->size[0] * originalImage->size[1] *
                                   plane];
        i = (int)(r + 1U);
        if (r + 1U > 255U) {
          i = 255;
        }
        loop_ub = (int)(r + 1U);
        if (r + 1U > 255U) {
          loop_ub = 255;
        }
        planeHist_data[i - 1] = planeHist_data[loop_ub - 1] + 1.0;
      }
    }
    loop_ub = originalHist_size[1];
    for (i = 0; i < loop_ub; i++) {
      originalHist_data[plane + 3 * i] = planeHist_data[i];
    }
  }
  return L;
}

/*
 * Arguments    : double L
 *                const double originalHist_data[]
 *                const emxArray_uint8_T *originalImage
 *                emxArray_uint8_T *equalizedImage
 * Return Type  : void
 */
static void equalize(double L, const double originalHist_data[],
                     const emxArray_uint8_T *originalImage,
                     emxArray_uint8_T *equalizedImage)
{
  double normalizer;
  double s;
  int M;
  int N;
  int i;
  int j;
  int plane;
  int x;
  int y;
  const unsigned char *originalImage_data;
  unsigned char r;
  unsigned char *equalizedImage_data;
  originalImage_data = originalImage->data;
  N = equalizedImage->size[0] * equalizedImage->size[1] *
      equalizedImage->size[2];
  equalizedImage->size[0] = originalImage->size[0];
  equalizedImage->size[1] = originalImage->size[1];
  equalizedImage->size[2] = 3;
  emxEnsureCapacity_uint8_T(equalizedImage, N);
  equalizedImage_data = equalizedImage->data;
  N = originalImage->size[0];
  M = originalImage->size[1];
  normalizer = (L - 1.0) / ((double)originalImage->size[0] *
                            (double)originalImage->size[1]);
#pragma omp parallel for num_threads(omp_get_max_threads()) private(           \
        s, r, y, x, i, j)

  for (plane = 0; plane < 3; plane++) {
    for (y = 0; y < N; y++) {
      for (x = 0; x < M; x++) {
        r = originalImage_data[(y + originalImage->size[0] * x) +
                               originalImage->size[0] * originalImage->size[1] *
                                   plane];
        s = 0.0;
        i = r;
        for (j = 0; j <= i; j++) {
          s += originalHist_data[plane + 3 * j];
        }
        s *= normalizer;
        s = rt_roundd_snf(s);
        if (s < 256.0) {
          if (s >= 0.0) {
            r = (unsigned char)s;
          } else {
            r = 0U;
          }
        } else if (s >= 256.0) {
          r = MAX_uint8_T;
        } else {
          r = 0U;
        }
        equalizedImage_data[(y + equalizedImage->size[0] * x) +
                            equalizedImage->size[0] * equalizedImage->size[1] *
                                plane] = r;
      }
    }
  }
}

/*
 * Arguments    : double u
 * Return Type  : double
 */
static double rt_roundd_snf(double u)
{
  double y;
  if (fabs(u) < 4.503599627370496E+15) {
    if (u >= 0.5) {
      y = floor(u + 0.5);
    } else if (u > -0.5) {
      y = u * 0.0;
    } else {
      y = ceil(u - 0.5);
    }
  } else {
    y = u;
  }
  return y;
}

/*
 * equalizedImage = histequalize(originalImage)
 *  Histogram equalization (or linearization) for improving image contrast.
 *  Given an NxMx3 image, equalizes the histogram of each of the three image
 *  planes in order to improve image contrast.
 *
 * Arguments    : const emxArray_uint8_T *originalImage
 *                emxArray_uint8_T *equalizedImage
 * Return Type  : void
 */
void histequalize(const emxArray_uint8_T *originalImage,
                  emxArray_uint8_T *equalizedImage)
{
  double originalHist_data[768];
  double L;
  int originalHist_size[2];
  if (!isInitialized_histequalize) {
    histequalize_initialize();
  }
  L = computeHistogram(originalImage, originalHist_data, originalHist_size);
  equalize(L, originalHist_data, originalImage, equalizedImage);
}

/*
 * File trailer for histequalize.c
 *
 * [EOF]
 */