PIL mean of all non-transparent/black pixels in RGBA image

Question

I want to achieve the same effect as in: cv::mean for non black pixel

However I am using PIL and converting a PIL image to cv image and back is too much overhead.

I have tried using mean_color = ImageStat.Stat(img).mean to get the mean color. However, this will include all transparent pixels too. I would like to calculate the mean of all pixels that have an alpha value above 0. So the mean over all non-completely-transparent pixels.

I am trying to keep the code nice and quick as I have to process a bunch of files. I was hoping for some built-in PIL function to do this, but couldn't find any.

Answer 1

It might not be the cleanest solution, but I got it to work.

def mean(rgb, a):
    """
    Supply with an RGB PIL Image and Alpha Channel PIL Image.
    Calculates the mean over all non-fully-transparent pixels in rgb.
    """

    a_arr = np.array(a)       # Convert Alpha values Image to array.
    img_arr = np.array(rgb)   # Convert Image RGB values to array.
    mask = (a_arr > 0)        # Create mask from all non-transparent pixels
    stuff = img_arr[mask]     # Array containing all pixels that aren't transparent

    rows = len(stuff)         # Get the row size.
    if rows < 1:              # If all pixels are transparent:
        return (0, 0, 0)      # The mean is simply black
    cols = len(stuff[0])      # Else, continue with the size of cols

    data = np.zeros([cols, rows, 3], dtype = np.uint8) # Create an array to contain the pixels
    data[:] = stuff           # Put the pixels with at least a > 0 into the created array.

    c_img = Image.fromarray(data, 'RGB') # Convert back to RGB PIL Image
    return ImageStat.Stat(c_img).mean # Calculate the mean over all pixels

Performance-wise, it was enough for my case.

About 3.44 seconds to convert about a thousand 16x16 image files. The process was:

Taking the mean then saving a Image.new('RGB', (16, 16), mean) .