Dear Hossein,

there's lots of implementations of Haar feat. implementations out there, OpenCV for example. But if you insist in implementing it by yourself the proper way is through integral image. I suggest you to start by reading the original paper by Viola and Jones 2001.

https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf

Btw, I would suggest to start with square windows of M by M which is usually a good idea and then move to rectangular if you want.

Best,

Magí.

Hi Hossein.

Depending on what you want to do and if you are willing to use OpenCV (as suggested by Magi), this article may be useful: http://coding-robin.de/2013/07/22/train-your-own-opencv-haar-classifier.html

Cheers

Luiz

dear friends,

I have read most of OPENCV examples,

I want to implement these features in MATLAB but I have difficulties with understanding rectangle size, window size, how to apply this rectangles to sliding window(pixel by pixel or each rectangle for whole of sliding window) and ...

Dear Hossein,

Please read the slides in the link attached. There's some normalization you should apply to the patch when computing Haar. Not only by patch or window size, but also by light changes or absolute values of the wavelet. On slide 12 you'll find the final formula, but read all the slides if you are interested in why each formula part. Hope this solves your problem.

Best,

http://www.cvc.uab.es/adas/site/userfiles/files/haarfeatures.pdf

Thanks, Andorra

do you have any idea of rectangular size? or if you have a sliding window 24*24,how do you apply haar feature? can you give more details or a simple matlab code?

Thanks anyway

A Haar-like feature considers adjacent rectangular regions at a specific location in a detection window, sums up the pixel intensities in these regions and calculates the difference between them

it can be implemented as follows:

function prg_haar_features()

clc,close all,clear all

% parameters

hs = 20; % Haar filter size (integer and even)

fs = 10; % font size

imgPath = './images/.jpg'; % image file path

imgPath = './images/.jpg'; % image file path

% input image

img = imread(imgPath);

% image size

[sy,sx,nc] = size(img);

% show image

figure,imshow(img),title('Input Image','fontsize',fs),xlabel(sprintf('Size -> [%d x %d]',sy,sx),'fontsize',fs);

% convert to gray-scale image

if (nc==3), img = rgb2gray(img); end

% compute the integral image over the input image: img->II

tic; II = mex_img2II(double(img)); t1 = toc;

% compute the Haar-like features in the integral image (II)

tic; [Hx,Hy] = mex_haar_features(II,hs); t2 = toc;

% gradient magnitud: the Haar-like features are used as local

% derivative operators

mag = sqrt(Hx.^2 + Hy.^2);

% messages

fun_messages('times:','process');

fun_messages(sprintf('img -> II : %.5f [sec.]',t1),'information');

fun_messages(sprintf('II -> {Hx,Hy} : %.5f [sec.]',t2),'information');

% show image

figure,subplot(131),imagesc(Hx),colormap(gray),xlabel('Hx','fontsize',fs);

subplot(132),imagesc(Hy),colormap(gray),xlabel('Hy','fontsize',fs);

subplot(133),imagesc(mag),colormap(gray),xlabel('Magnitude','fontsize',fs);

figure,imagesc(mag),colormap(hot),xlabel('Magnitude','fontsize',fs);

end

MEX FILE FOR HAAR FEATURES

/* mex_haar_features

Description:

This program computes Haar-like features over a given input image (img) in

order to extract contours. The Haar-like features are used as local derivative

operators. Particularly, the program computes horizontal (Hx) and vertical (Hy)

oriented features (Haar-like features) using the integral image (II) of the

input image (img). This initial pre-processing step is done in advance using a

mex file (mex_img2II.cc).

Input:

prhs[0] Haar filter response (Hy)

Contact:

Michael Villamizar

[email protected]

Institut de Robòtica i Informàtica Industrial CSIC-UPC

Barcelona - Spain

2014

*/

#include

#include "mex.h"

#include

// main function

void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {

// check

if(nrhs!=2)

mexErrMsgTxt("Two inputs required: 1. input integral image (II) 2. Haar filter size");

if(nlhs!=2)

mexErrMsgTxt("Two outputs required: 1. Haar filter response (Hx) 2. Haar filter response (Hy)");

MEX FILE for Integral image

/* mex_img2II

Description:

This function computes the integral image (II) over the input image (img).

Input:

prhs[0] integral image (II)

Contact:

Michael Villamizar

[email protected]

Institut de Robòtica i Informàtica Industrial CSIC-UPC

Barcelona - Spain

2014

*/

#include

#include "mex.h"

#include

// main function

void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {

// check

if(nrhs!=1)

mexErrMsgTxt("One input required: input image (img)");

if(nlhs!=1)

mexErrMsgTxt("One output required: integral image (II)");

// input image

double *img = (double *)mxGetPr(prhs[0]); // image data

const int *imgSize = mxGetDimensions(prhs[0]); // image dimensions

// image size

int sy = imgSize[0];

int sx = imgSize[1];

// num. image dimensions

int nd = mxGetNumberOfDimensions(prhs[0]);

// num. image channels

int nc = 1; // gray-scale image

if (nd==3){ nc = imgSize[2];} // color image

// output data: integral image

int out[3] = {sy,sx,nc};

plhs[0] = mxCreateNumericArray(3, out, mxDOUBLE_CLASS, mxREAL);

double *II = (double *)mxGetPr(plhs[0]);

// variables

double value,xo,yo,xoyo;

// integral image computation -recursive computation-

for (int c=0; c0) {

xo = 0;

yo = *(II + (y-1) + x*sy + c*sy*sx);

xoyo = 0;

}

else {

xo = 0;

yo = 0;

xoyo = 0;

}

}

// integral image value at (x,y,c)

*(II + y + x*sy + c*sy*sx) = value - xoyo + xo + yo;

}

}

}

}

// integral image (II)

double *II = (double *)mxGetPr(prhs[0]);

const int *imgSize = mxGetDimensions(prhs[0]);

// Haar filter size

int hs = (int)mxGetScalar(prhs[1]);

// image size

int sy = imgSize[0];

int sx = imgSize[1];

// middle point

int mp = (int)round((double)hs/2);

// output maps limits

int ly = sy-hs-1;

int lx = sx-hs-1;

// output feature maps (Hx, Hy)

int out[2] = {ly,lx};

plhs[0] = mxCreateNumericArray(2, out, mxDOUBLE_CLASS, mxREAL);

plhs[1] = mxCreateNumericArray(2, out, mxDOUBLE_CLASS, mxREAL);

double *HxMap = (double *)mxGetPr(plhs[0]); // Hx map

double *HyMap = (double *)mxGetPr(plhs[1]); // Hy map

// variables

int x1,y1,x2,y2;

double area1,area2,Hx,Hy;

// sliding window

for (int x=0; x

so It seems as if openCV doesnt allow you to shape your own haar filter