Opencv image sharpness evaluation (camera autofocus)

Recently, I want to realize camera auto focusing , That is, image sharpness evaluation is required , Search on the Internet .

The first good article is ： Sharpness evaluation method without reference image
 http://nkwavelet.blog.163.com/blog/static/227756038201461532247117

This is a very detailed introduction to the theory or formula , share 16 A calculation method , It's worth learning .

（1）Brenner Gradient function

（2）Tenengrad  Gradient function

（3）Laplacian  Gradient function

（4）SMD（ Gray variance ） function

...

Next is fuzzy detection , Noise detection , Combination of noise and blur

The article pays attention to theory , There is no specific sample program . There is a result comparison chart .

The second good article is ：OpenCV Image sharpness evaluation （ Camera autofocus ）
https://blog.csdn.net/dcrmg/article/details/53543341

The implementation of the first three methods in this set , this 3 It's in opencv All have function implementation . Realization 3 Three kinds of sharpness evaluation methods , Namely Tenengrad Gradient method 、Laplacian Gradient method and variance method .

Tenengrad Gradient method , The measurement index is through Sobel The average gray value of the image processed by the operator , The bigger the value is. , Represents the sharper the image .Sobel function .

Laplacian Gradient method ： ditto , But with  Laplacian function

Variance method ： Variance is used in probability theory to examine a set of discrete data and its expectations （ That is, the mean value of the data ） The separation between （ deviate ） Measurement method of degree . The variance is large , It means that the deviation between this group of data is large , Some of the data in the group are large , Some are smaller , It's not evenly distributed ; Variance is small , It means that the deviation between this group of data is small , The data in the group is evenly distributed , Similar in size . The greater the variance , It means the better the clarity .

This sample program is available for reference .

The code of the first algorithm is as follows （ The second is to write off some differences ）：

#include <highgui/highgui.hpp>
#include <imgproc/imgproc.hpp>
 
using namespace std;
using namespace cv;
 
int main()
{
    Mat imageSource = imread("2.jpg");
    Mat imageGrey;
 
    cvtColor(imageSource, imageGrey, CV_RGB2GRAY);
    Mat imageSobel;
    Sobel(imageGrey, imageSobel, CV_16U, 1, 1);
  //Laplacian(imageGrey, imageSobel, CV_16U); Algorithm 2
 
    // The average gray level of the image
    double meanValue = 0.0;
    meanValue = mean(imageSobel)[0];
 
    //double to string
    stringstream meanValueStream;
    string meanValueString;
    meanValueStream << meanValue;
    meanValueStream >> meanValueString;
    meanValueString = "Articulation(Sobel Method): " + meanValueString;
    putText(imageSource, meanValueString, Point(20, 50), CV_FONT_HERSHEY_COMPLEX, 0.8, Scalar(255, 255, 25), 2);
    imshow("Articulation", imageSource);
    waitKey();
}
The evaluation value here is meanValue , The larger the value, the clearer .// double to string Part is just added to the figure to show .

The difference in variance calculation is ：

// Find the standard deviation of gray image
    meanStdDev(imageGrey, meanValueImage, meanStdValueImage);
    double meanValue = 0.0;
    meanValue = meanStdValueImage.at<double>(0, 0);
It is also mentioned that ：

In industrial applications , The clearest focusing image is not necessarily the best , There may be molars （ Water ripple ） The phenomenon , Generally, you need to make a fine adjustment in the attachment of the clearest focusing position .