How to optimize this image iteration in numpy?

问题

I'm using this code to detect green color in the image.

The problem is this iteration is really slow.

How to make it faster? If it is using numpy, How to do it in numpy way?

def convertGreen(rawimg):
    width, height, channels = rawimg.shape
    size = (w, h, channels) = (width, height, 1)
    processedimg = np.zeros(size, np.uint8)
    for wimg in range(0,width):
        for himg in range(0,height):
            blue = rawimg.item(wimg,himg,0)
            green = rawimg.item(wimg,himg,1)
            red = rawimg.item(wimg,himg,2)
            exg = 2*green-red-blue
            if(exg > 50):
                processedimg.itemset((wimg,himg,0),exg)

    return processedimg

回答1:

Try simply this:

blue = rawimg[:,:,0]
green = rawimg[:,:,1]
red = rawimg[:,:,2]
exg = 2*green-red-blue
processedimg = np.where(exg > 50, exg, 0)

回答2:

I've only dabbled with numpy as a hobbyist, but I believe that you could take advantage of fromfunction which creates a new np array from an existing one https://docs.scipy.org/doc/numpy/reference/generated/numpy.fromfunction.html

Here is what I think might work in that case - which would take advantage of numpy's speed:

def handle_colors(img, x, y):
    blue = img.item(x,y,0)
    green = img.item(x,y,1)
    red = img.item(x,y,2)
    exg = 2*green-red-blue
    if exg > 50:
        return (exg, green, red)
    return blue, green, red

def convertGreen(rawimg):
    processedimg = np.fromfunction(lambda i, j: handle_colors(rawimg, i, j), rawimg.shape)
    return processedimg

来源：https://stackoverflow.com/questions/42991524/how-to-optimize-this-image-iteration-in-numpy