更快的图像归一化（numpy 数组）

Question

I've got function that takes some image as a input and changes it values (scales) in a way that average value will be 96 . Here is the function:

def normalize_image(image: np.ndarray):
    image_median = np.median(image[image > 0])
    image = image * 96.0 / image_median
    image[image > 255] = 255
    return image

I am using python 3.5.3 and numpy 1.15.2. I profiled my code with cProfile and it turned out that this function takes 6% of all time (in some scenarios up to 25% of all time) having only 50 calls. These array have shape of (155,256,256).

I am not very experienced with optimizing python and I wonder if this could be made faster somehow?

Normally I would start with using SIMD optimization but numpy use them heavily already.

Answer 1

There is not much you can do quickly here.

You already use NumPy in a vectorized manner, so internally C code is executed that probably is already quite optimized.

The calculation of the median can take much longer than calculating the mean (because sorting is involved). Consider replacing it.

Adding some parentheses should save a division of the array

image = image * (96.0 / image_median)

because between operators of equal precedence Python goes from left to right.