How to make Gaussian Blur work properly?

Question

I'm writing a Gaussian Blur filter with 3x3 matrix(kernel) and using QImage as external library. The kernel is calculated as it has to be. After kernel is initialized and calculated I bring the unsigned int Matrix into the program:

typedef std::vector<std::vector<unsigned int> > uintMatrix; // For copying pixels value

In the main.cpp it's initialized like:

uintMatrix tmpPxMat(image.width(); std::vector<unsigned int> (image.height(), 0));

I use it as a temporary storage for blurred pixels values.

Then, I use my algorithm to blur the image:

// Blur Pixels
for(int x = 0; x < image.width(); x++)
    for(int y = 0; y < image.height(); y++) {
        // To avoid edge cases don't process them, I'll fix it soon
        if(x == 0 || x == image.height()-1 || y == 0 || y == image.width()-1)
            continue;
        // 3x3 Matrix
        tmpPxMat[x][y] += kernel[0][0] * image.pixel(x-1, y-1) +
                          kernel[0][1] * image.pixel(x, y-1) +
                          kernel[0][2] * image.pixel(x+1, y-1);

        tmpPxMat[x][y] += kernel[1][0] * image.pixel(x-1, y) +
                          kernel[1][1] * image.pixel(x, y) +
                          kernel[1][2] * image.pixel(x+1, y);

        tmpPxMat[x][y] += kernel[2][0] * image.pixel(x-1, y+1) +
                          kernel[2][1] * image.pixel(x, y+1) +
                          kernel[2][2] * image.pixel(x+1, y+1);
    }

Then, I copy the result of tmpPxMat[][] to the origin image:

// Copy blurred values to the image
for(int x = 0; x < image.width(); x++)
        for(int y = 0; y < image.height(); y++)
            image.setPixel(x, y, tmpPxMat[x][y]);

And save it:

image.save("/home/john/Pictures/blurred", "PNG", 100);

But at the end I get not the result I was waiting for. Here what I get:

Before / After:

Origin/Blurred

Sorry for the long question description, but I compressed it as much as I could.

Are you sure there is no problem with casting? Gaussian kernel should be a float matrix, the result of your convolution being a float saved as an int. — Miki, Jul 01 '15 at 14:21
Your output image is coloured, yet your processing doesn't seem to be processing three channels of colour data (RGB) so are you using greyscale processing on colour images? — Mark Setchell, Jul 01 '15 at 14:49
@Miki, my kernel is float matrix, but temporary storage is not. I've changed it to float as needed, but at the end - I use **setPixel()** function, which accepts **uint**, I can avoid casting when calculate values, but can't when setting these values to the pixel. The casting should be very small, can it damage picture so hard? — pushandpop, Jul 01 '15 at 14:53
It was just a guess.. probably wrong! @MarkSetchell question seems more appropriate — Miki, Jul 01 '15 at 15:15
@Miki, I think he's 99% right. I'm trying now to convert it to RGB32 or load it properly or whatever. I'll leave feedback. — pushandpop, Jul 01 '15 at 15:22
@MarkSetchell, I've just found this question on [SO](http://stackoverflow.com/questions/25216834/converting-2d-mask-to-1d-in-gaussian-blur?rq=1) and the OP was having the same effect on the resulting image. As one of the commentators said: "2d kernel must have total sum of 1, otherwise it won't work". I traced my ``sum`` variable and it equals ~4.xxxxx which is not right. I'm gonna try fix it right now. — pushandpop, Jul 01 '15 at 15:51
@MarkSetchell, Oh, no. I was checking the total sum of kernel before normalizing it. I've just checked it after normalizing and it equals to ``1``. So, the problem is not in the kernel. I should go deeper and find mistake somewhere else. — pushandpop, Jul 01 '15 at 15:58

score 3 · Accepted Answer · answered Jul 01 '15 at 16:07

3

I assume uintMatrix is a two-dimensional array of 32-bit ints, and that you've packed the red, green, and blue channels into that.

If so, that's your problem. You need to blur each channel independently.

answered Jul 01 '15 at 16:07

Adrian McCarthy

45,555
16
123
175

Yes, you're exactly right. It works like a charm. So, you're my hero today, I was struggling with implementation of this filter for a few days. Does the strength of the blur depend on a size of a kernel? If I increase it to 5x5, will it be more blured? – pushandpop Jul 01 '15 at 16:50
1

@pushandpop - yes it will. The larger the size the larger the blur. Experiment with different sizes and see for yourself! – rayryeng Jul 01 '15 at 17:16

How to make Gaussian Blur work properly?

1 Answers1