I wanted to try my hand at text recognition, so i've used opencv to trace out the edges and c++ to find slopes, curves etc, the edge algorithm works well on big and uncluttered sets of characters but when it comes against small printed text or text with a lot of background noise like embedded in captcha it struggles and looks incomplete, my guess was i hadn't set the threshold values correctly and tried different values with no success.
Here is my code :
#include "cv.h"
#include "highgui.h"
using namespace cv;
const int low_threshold = 50;
const int high_threshold = 150;
int main()
{
IplImage* newImg;
IplImage* grayImg;
IplImage* cannyImg;
newImg = cvLoadImage("ocv.bmp",1);
grayImg = cvCreateImage( cvSize(newImg->width, newImg->height), IPL_DEPTH_8U, 1 );
cvCvtColor( newImg, grayImg, CV_BGR2GRAY );
cannyImg = cvCreateImage(cvGetSize(newImg), IPL_DEPTH_8U, 1);
cvCanny(grayImg, cannyImg, low_threshold, high_threshold, 3);
cvNamedWindow ("Source", 1);
cvNamedWindow ("Destination",1);
cvShowImage ("Source", newImg );
cvShowImage ("Destination", cannyImg );
cvWaitKey(0);
cvDestroyWindow ("Source" );
cvDestroyWindow ("Destination" );
cvReleaseImage (&newImg );
cvReleaseImage (&grayImg );
cvReleaseImage (&cannyImg );
return 0;
}
I've looked across the net and have seen some complicated thresholding conditions like in this code from this site :
% Set direction to either 0, 45, -45 or 90 depending on angle.
[x,y]=size(f1);
for i=1:x-1,
for j=1:y-1,
if ((gradAngle(i,j)>67.5 && gradAngle(i,j)<=90) || (gradAngle(i,j)>=-90 && gradAngle(i,j)<=-67.5))
gradDirection(i,j)=0;
elseif ((gradAngle(i,j)>22.5 && gradAngle(i,j)<=67.5))
gradDirection(i,j)=45;
elseif ((gradAngle(i,j)>-22.5 && gradAngle(i,j)<=22.5))
gradDirection(i,j)=90;
elseif ((gradAngle(i,j)>-67.5 && gradAngle(i,j)<=-22.5))
gradDirection(i,j)=-45;
end
end
end
If this is the solution can somebody provide me the c++ equivalent of this algorithm, if it's not what else can i do ?
Canny edge detector is a multi-step detector using hysteresis thresholding (it uses two threshold instead of one), and edge tracking (your last snippet is the part of this step). I suggest reading the wikipedia entry first. One possible solution could be to choose the high threshold, so e.g. 70% of the image pixels would be classified as edge (initially - you could do this quickly using histograms), than choose the low threshold as e.g. 40% of the high threshold. It might be a good idea to try to perform edge detection on image block rather than the whole image, so your algorithm could calculate different thresholds for different areas.
Note that CAPTCHA-s are designed to be hard to segment, and adding noise that broke edge detection is one technique to achive this (you might need to smooth the image first).