Opencv slow code: Is something wrong?

Question

here is a function where I try to improve the image colors. It works but it is really slow...Maybe someone has a better idea?

static Mat correctColor(Mat AImage) {
    Mat copyImage;
    AImage.copyTo(copyImage);
    Mat imgLab;
    cvtColor(copyImage, imgLab, CV_BGR2Lab);


    for (int y = 0; y < imgLab.rows; y++) {
        for (int x = 0; x < imgLab.cols; x++) {
            //get pixel value                
            imgLab.ptr<uchar > (y)[x * 3] = imgLab.ptr<uchar > (y)[x * 3]*0.3;
        }
    }
    cvtColor(imgLab, copyImage, CV_Lab2BGR);

    Mat img(copyImage.rows, copyImage.cols, CV_32FC3);
    copyImage.convertTo(img, CV_32FC3, 1 / 255.);        
    blur(img, img, Size(255, 255));
    Mat img32(copyImage.rows, copyImage.cols, CV_32FC3);
    copyImage.convertTo(img32, CV_32FC3, 1 / 255.);

    img.mul(img, 2);        
    divide(img32, img, img);

    img.convertTo(copyImage, CV_8U, 255.0);

    return copyImage;
 }

This should be better posted to Code Review

CharlesB
– CharlesB

2012-05-07 12:19:08 +00:00
Commented May 7, 2012 at 12:19 — CharlesB
– CharlesB, Commented May 7, 2012 at 12:19

Eric · Accepted Answer · 2012-05-07 14:39:38Z

2

The major problem is that you are creating several copies of the original image in memory: AImage, copyImage, imgLab, img, img32.

First optimization should be what @Eric suggested (pass by reference):

static Mat correctColor(Mat& AImage) {

As for the rest of your code, see if you can decrease the number of copies you work with.

OpenCV has a GPU module which implements several functions in the GPU, including cv::blur(). This implementation is based on the CUDA framework, so if your graphics card is NVIDIA you are in luck: gpu::blur().

edited May 7, 2012 at 14:39

Eric

2,3413 gold badges23 silver badges31 bronze badges

answered May 7, 2012 at 12:47

karlphillip

93.7k38 gold badges254 silver badges448 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

Sam · Accepted Answer · 2012-05-07 17:21:12Z

2

The best way to optimize is to start where you spend the most time. So, I strongly recommend you to profile this code to know exactly which parts of your code are the most time-consuming.

Now, some general ideas on how to improve:

If there is any chance to get rid of the convert to Lab, do it.
Do not use floats. There are a number of other ways to do math, without floating point. So, try first on uchars (get rid of the normalization step), then, if the approximation is too bad, convert to unsigned short, then back to uchar, when all the calculations are done.
Try to find another operation instead of divide. It is one of the most expensive math operations you can do on a computer. Maybe you can shift bits, or find an operation with a similar result.
Do not access elements by ptr. Search on SO for fast Mat access.
blur on a 255x255 window is a killer. I suspect it takes 95% of your processing time. You really cannot make it smaller? or use a box filter? or not use it at all? Or look on the internet for a FFT filter implementation. It is much faster than the simple window approach used by opencv.

answered May 7, 2012 at 17:21

Sam

20.1k4 gold badges64 silver badges85 bronze badges

2 Comments

solvingPuzzles Over a year ago

Why not access elements by ptr? @vasile

Sam Over a year ago

accessing elements by image.ptr<uchar> is slower than raw pointer access (extract a raw pointer to the image data, and access it by pointer indexes)

Eric · Accepted Answer · 2012-05-07 12:15:34Z

1

First of all, you should pass your argument by reference since you already create a clone in your code.

answered May 7, 2012 at 12:15

Eric

2,3413 gold badges23 silver badges31 bronze badges

Comments

Federico Cristina · Accepted Answer · 2012-05-07 12:31:20Z

0

Improve image colors? You should try Histogram Equalization instead. Look for the equalizeHist() function.

edited May 7, 2012 at 12:31

answered May 7, 2012 at 12:20

Federico Cristina

2,2431 gold badge19 silver badges38 bronze badges

2 Comments

Sebastian Mach Over a year ago

That's not a panacea of "improvement". Improvement is not a mathematically defined term ;)

Federico Cristina Over a year ago

Well... you are right. :) But equalization is very useful for these kind of tasks.

Simon West · Accepted Answer · 2015-12-09 09:04:04Z

-2

Apart from the optimizations to the programs. You can add compiler optimization flags like -o3 and -NDEBUG, while compiling.

edited Dec 9, 2015 at 9:04

Simon West

3,7881 gold badge28 silver badges28 bronze badges

answered Dec 9, 2015 at 8:37

Santhosh Nagasanthosh

231 silver badge2 bronze badges

Collectives™ on Stack Overflow

Opencv slow code: Is something wrong?

5 Answers 5

Comments

2 Comments

Comments

2 Comments

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

5 Answers 5

Comments

2 Comments

Comments

2 Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related