Image edge smoothing with opencv

Question

I am trying to smooth output image edges using opencv framework, I am trying following steps. Steps took from here https://mcmap.net/q/590091/-how-to-refine-or-blur-or-smooth-just-the-edges

int lowThreshold = 10.0;
int ratio = 3;
int kernel_size = 3;

Mat src_gray,detected_edges,dst,blurred;

/// Convert the image to grayscale
cvtColor( result, src_gray, CV_BGR2GRAY );

/// Reduce noise with a kernel 3x3
cv::blur( src_gray, detected_edges, cv::Size(5,5) );

/// Canny detector
cv::Canny( detected_edges, detected_edges, lowThreshold, lowThreshold*ratio, kernel_size );

//Works fine upto here I am getting perfect edge mask    

cv::dilate(detected_edges, blurred, result);

//I get Assertion failed (src.channels() == 1 && func != 0) in countNonZero ERROR while doing dilate

result.copyTo(blurred, blurred);

cv::blur(blurred, blurred, cv::Size(3.0,3.0));

blurred.copyTo(result, detected_edges);

UIImage *image = [UIImageCVMatConverter UIImageFromCVMat:result];

I want help whether if I am going in right way, or what am I missing?

Thanks for any suggestion and help.

Updated:

I have got an image like below got from grabcut algorithm, now I want to apply edge smoothening to the image, as you can see the image is not smooth.

Hilltop

@Boyko Perfanov I need help in it. I am trying to follow the steps you stated here https://mcmap.net/q/590091/-how-to-refine-or-blur-or-smooth-just-the-edges – Hilltop

Answer 1

Do you want to get something like this?

If yes, then here is the code:

#include <iostream>
#include <vector>
#include <string>
#include <fstream>
#include <opencv2/opencv.hpp>

using namespace cv;
using namespace std;

int main(int argc, char **argv)
{
    cv::namedWindow("result");
    Mat img=imread("TestImg.png");
    Mat whole_image=imread("D:\\ImagesForTest\\lena.jpg");
    whole_image.convertTo(whole_image,CV_32FC3,1.0/255.0);
    cv::resize(whole_image,whole_image,img.size());
    img.convertTo(img,CV_32FC3,1.0/255.0);
    
    Mat bg=Mat(img.size(),CV_32FC3);
    bg=Scalar(1.0,1.0,1.0);

    // Prepare mask
    Mat mask;
    Mat img_gray;
    cv::cvtColor(img,img_gray,cv::COLOR_BGR2GRAY);
    img_gray.convertTo(mask,CV_32FC1);
    threshold(1.0-mask,mask,0.9,1.0,cv::THRESH_BINARY_INV);

    cv::GaussianBlur(mask,mask,Size(21,21),11.0);
    imshow("result",mask);
    cv::waitKey(0);


        // Reget the image fragment with smoothed mask
    Mat res;

    vector<Mat> ch_img(3);
    vector<Mat> ch_bg(3);
    cv::split(whole_image,ch_img);
    cv::split(bg,ch_bg);
    ch_img[0]=ch_img[0].mul(mask)+ch_bg[0].mul(1.0-mask);
    ch_img[1]=ch_img[1].mul(mask)+ch_bg[1].mul(1.0-mask);
    ch_img[2]=ch_img[2].mul(mask)+ch_bg[2].mul(1.0-mask);
    cv::merge(ch_img,res);
    cv::merge(ch_bg,bg);

    imshow("result",res);
    cv::waitKey(0);
    cv::destroyAllWindows();
}

And I think this link will be interestiong for you too: Poisson Blending

Octopod

Thanks for the answer, but I don't want output like this, I have updated my question, see if you can help please, thanks. – Hilltop

I've changed my answer. Hope you'll catch the idea. – Octopod

WHen I try to implement the above logic, I get something like this i.imgur.com/ClXXpbQ.png?1, it seems its doing things correctly but output is very different.. Any Idea.? – Hilltop

It looks like a problem with CV_8UC3/CV_8UC1 and CV_32FC3/CV_32FC1 image types. The range for the first 0-255, range for the second one is 0-1. Try convert images you work with to CV_32FC3/CV_32FC1 type and scale it by factor 1.0/255.0 . I suspect that the values in mask image you used are in 0-255 range instead of 0-1. – Octopod

Also I should mention that used images are: TestImage.jpg - is the image you povide above, lena.jpg - is initial image where you got TestImage.jpg from. – Octopod

Yes exactly I am doing the same, the testimage in my case is the grabcut result I am getting, and lena is the original image. – Hilltop

And what about image types and pixel value ranges? – Octopod

Image type is CV_8UC1 not sure about pixel value ranges – Hilltop

You should convert it to CV_32FC1. And scale by factor 1.0/256.0 . All images should have real element type. I've done these conversions in the code above. – Octopod

let us continue this discussion in chat – Hilltop

I am unable to get the output image, and can you please help me ? – Hilltop

Can I see your code? You can send it to my email: andreysmorodovatlistdotru. Replace "at" and "dot" with "@" and ".". – Octopod

@Hilltop Did you get something for edge smoothing? – Amused

How do you save the result as a .bmp file ? I tried it but I get a very dark image. – Radmen

Just multiply it by 255.In float format it is in 0-1 range. – Octopod

can i get java solution of this answer? – Supervisor

@AndreySmorodov"Poisson Blending" link has been broken. – Supervisor

@Hilltop can you convert that code in java,it will be very helpful. – Supervisor

Thanks! Link retargeted. Anyway you can find larger amount of papers and implementations on "poisson blending" or "compositing" keywords. – Octopod

Answer 2

I have followed the following steps to smooth the edges of the Foreground I got from GrabCut.

1. Create a binary image from the mask I got from GrabCut.
2. Find the contour of the binary image.
3. Create an Edge Mask by drawing the contour points. It gives the boundary edges of the Foreground image I got from GrabCut.
4. Then follow the steps define in https://mcmap.net/q/590091/-how-to-refine-or-blur-or-smooth-just-the-edges

Revell