CODEWITHSUNDEEP
pythonperformanceopencvnumpy
BFlint
BFlint

Reputation: 2827

How to increase pixel math speed using NumPy

I'm looking for help on how to increase the speed of this calculation. What I'm trying to do is access each pixel and do some math on it, then create a new image with the new pixel calculations. I'm running this through a few thousands of small images which takes 1hr+. Any help would be appreciated, thanks.

image=cv2.imread('image.png')

height, width, depth = image.shape

for i in range(0, height):  
    for j in range (0, width):
        B = float(image.item(i,j,0)) #blue channel of image
        R=float(image.item(i,j,2)) #red channel of image

        num = R-B
        den = R+B

        if den == 0:
            NEW=1
        else:
            NEW = ((num/den)*255.0)

        NEW = min(NEW,255.0)
        NEW = max(NEW,0.0)
        image[i,j] = NEW  #Sets all BGR channels to NEW value

cv2.imwrite('newImage.png',image)

Upvotes: 0

Views: 252

Answers (1)

unutbu
unutbu

Reputation: 879351

Remove the double for-loop. The key to speed with NumPy is to operate on the whole array at once:

image = cv2.imread('image.png')    
height, width, depth = image.shape

image = image.astype('float')
B, G, R = image[:, :, 0], image[:, :, 1], image[:, :, 2]
num = R - B
den = R + B
image = np.where(den == 0, 1, (num/den)*255.0).clip(0.0, 255.0)

cv2.imwrite('newImage.png',image)

By calling NumPy functions on whole arrays (rather than doing Python operations on scalar pixel values), you off-load most of the computational work to fast C/C++/Cython (or Fortran) compiled code called by the NumPy functions.

Upvotes: 4

Related Questions