What can I do for eliminate a "reflection light" of a image? using OpenCV

Question

My imagem has a "light reflection", the two first zeros on the image has some light different of the rest of the image. Whe I convert this to a binary image, this part becomes white, and I need to get the exact contour of the number and this hinders. ow I could solve this by using OpenCV?

the original image https://docs.google.com/file/d/0BzUNc6BOkYrNNlE3U04wWEVvVE0/edit?usp=sharing

the binary version https://docs.google.com/file/d/0BzUNc6BOkYrNeEE0U3NvOElqa1E/edit?usp=sharing

If I increase the value of the threshold, I lose the numbers on the right side of the image. My code:

#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/highgui/highgui.hpp>

using namespace cv;

int main ( int argc, char **argv )
{
   Mat im_gray = imread("img2.jpg",CV_LOAD_IMAGE_GRAYSCALE);

   Mat im_rgb  = imread("img2.jpg");
   cvtColor(im_rgb,im_gray,CV_RGB2GRAY);

   Mat img_bw = im_gray > 90;

   imwrite("image_bw2.jpg", img_bw);

   return 0;
}  

Answer 1

Shadows and glares are not easy problems to work with. But with some good work, they are possible to overcome.

Another step is to use your thresholded image as a mask to get another thresholded image. Here are some criteria that have worked for me:

• Restricting all but the dominant peak in the pixels contained in the histogram of the intermediate (what you have right now) thresholded image
• Use the derivative to find boundaries (cvSobel may help)
• Use a combination of strict adaptive and liberal hard thresholding to take into account the varying illumination of different parts of the image

Answer 2

Actually problem is not so difficult in your case. Because you have just 10 different numbers, trains some classifier to recognize them.

For fast start you can use http://blog.damiles.com/2008/11/basic-ocr-in-opencv.html

It will work because defects also repeat to some extent. You can train algorithm to recognize images with defects and forget about removing these.