how to get the result in template matching code? - ios

Question

I am a beginner to Computer vision .I am currently working on a project to find the match between two images using matchTemplate in iOS.The problem that I am facing is with finding a way to determine whether the two images are matching or not although matchTemplate is working well.I thought of taking the percentage of result matrix but I did not know how and could not find a way.also MinMaxLoc did not work with me .
If anyone can help me or give me an idea I would really really appreciate it cause I am on desperate point now.
Here is the code:
`
UIImage* image1 = [UIImage imageNamed:#"1.png"];
UIImage* image2 = [UIImage imageNamed:#"Image002.png"];
// Convert UIImage* to cv::Mat
UIImageToMat(image1, MatImage1);
UIImageToMat(image2, MatImage2);
MatImage1.resize(100 , 180);
MatImage2.resize(100 , 180);
if (!MatImage1.empty())
{
// Convert the image to grayscale
//we can also use BGRA2GRAY : Blue , Green , Red and Alpha(Opacity)
cv::cvtColor(MatImage1, grayImage1, cv::COLOR_BGRA2GRAY );
cv::cvtColor(MatImage2, grayImage2, cv::COLOR_BGRA2GRAY);
}
/// Create the result matrix
int result_cols = grayImage1.cols ;
int result_rows = grayImage1.rows ;
result.create( result_cols, result_rows, CV_32FC1 );
/// Do the Matching and Normalize
matchTemplate( grayImage1 , grayImage2 , result , CV_TM_SQDIFF_NORMED);
//Normalize
normalize( result, result, 0, 100, cv::NORM_MINMAX, -1 );
//Threshold
cv::threshold(result , result , 30, 0, CV_THRESH_TOZERO);`

Answer 1

The intent of matchTemplate(...) is that the template is usually smaller than the image. The template is then moved across the image as a sliding window and a 'matching score' is calculated in some way e.g. using cross-correlation or squared difference.
So if, the input image is 10x10 and the template is 3x3, then the template is positioned so the top left corner is at the top left corner of the image (centre of template is at pixel (1,1) assuming that we index from 0). The matching score is calculated and then the template slides to (1,2) and we match again. When the template's middle pixel is at (8,1) we slide it down to the next row (1,2) and repeat.
The output result of this process is an 8x8 matrix where the value at each position represents the matching score for when the template was at that point. The size of the output image is W-w+1 x H-h+1 where WxH is the size of the image and wxh is the size of the template.
You can then use minMaxLoc to work out which is the highest and lowest scores in the output matrix and, depending on the matching score you use, one of these will be the most likely match for the template within the image.
Now you are resizing your template and image to the same size:
MatImage1.resize(100 , 180);
MatImage2.resize(100 , 180);
which means that there is only one place that the template can be located within the image and your output matrix should be a 1x1 grid.
You are also using
CV_TM_SQDIFF_NORMED
Which is the normalised squared difference. For this score, a lower matching score is better. i.e. the closer the value in your 1x1 output matrix is to 0, the closer the match between your template and your image.
Given the sizes of template and image are 100x180 then you can easily calculate the maximum value that this matching score can have as 100x180x255 which you would get if the entire image were black and the template were white or vice-versa. This should help you work out a sensible threshold below which you would say t=your template matched the image.
Since you only have a 1x1 output though there's little value in normalising or thresholding the result.