How to recognize digits from the analog counter? - opencv

Question

I'm trying to read the following kWh numbers from the counter. The problem is the tesseract OCR doesn't recognize the analog digits.
The question is: will it be a better idea to make the photos of all of the digits (from 0 to 9) at different positions (I mean when digit is in the center, when it is a little at the top and the number 2 is appearing etc.) and to try image recognition instead of text recognition?
As far as I understood the difference is, that the image recognition compares the photos, while the text recognition... well I don't know...
Any advice?

Answer 1

Since the counter is not digital, but analog, we have problems at the transitions. The text/number recognition libraries can not recognize smth like that. The solution, that I've found is: Machine Learning.
Firstly I've made user to make the picture, where the numbers take 70-80% of the image (in order to remove the unneeded details).
Then I'm looking for parallel lines (if there are any) and cut the picture, that is between them (if the distance is big enough).
After that I'm filtering the picture (playing with contrast, brightness, set it grayscale) and then use the filter, that makes the image to contain only two colours (#000000 (black) and #ffffff (white)). In order to find the contours easier.
Then I find the contours by using Canny algorithm and filter them, by removing the unneeded details.
After that I use K-Nearest-Neighbour algorithm in order to recognize the digits.
But before I can recognize anything, I need to teach the algorithm, how the digits look like and what are they.
I hope it was useful!

Answer 2

Maybe you are not configuring tesseract right. I made a code using it that solves your problem:
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <tesseract/baseapi.h>
#include <iostream>
using namespace cv;
int main(int argc, char** argv)
{
cv::Mat input = cv::imread("img.jpg");
//rectangle containing just the kWh numbers
Rect roi(358,327,532,89);
//convert to gray scale
Mat input_gray;
cvtColor(input(roi),input_gray,CV_BGR2GRAY);
//threshold image
Mat binary_img = input_gray>200;
//make a copy to use on findcontours
Mat copy_binary_img = binary_img.clone();
vector<vector<Point> > contours;
vector<Vec4i> hierarchy;
//identify each blob in order to eliminate the small ones
findContours(copy_binary_img, contours, hierarchy, CV_RETR_TREE, CV_CHAIN_APPROX_SIMPLE, Point(0,0));
//filter blobs by their sizes
for (vector<vector<Point> >::iterator it = contours.begin(); it!=contours.end(); )
{
if (it->size()>20)
it=contours.erase(it);
else
++it;
}
//Erase blobs which have countour size smaller than 20
for( int i = 0; i< contours.size(); i++ )
{
drawContours( binary_img, contours, i, 0, -1, 8, hierarchy, 0, Point() );
}
//initialize tesseract OCR
tesseract::TessBaseAPI tess;
tess.Init(NULL, "eng", tesseract::OEM_DEFAULT);
tess.SetVariable("tessedit_char_whitelist", "0123456789-.");
tess.SetPageSegMode(tesseract::PSM_SINGLE_BLOCK);
//set input
tess.SetImage((uchar*)binary_img.data
, binary_img.cols
, binary_img.rows
, 1
, binary_img.cols);
// Get the text
char* out = tess.GetUTF8Text();
std::cout << out << std::endl;
waitKey();
return 0;
}