Compute mean and standard deviation (std) per channel give a list of images

compute_mean_std_per_channel.py

"""Give a list of image filenames, this script compute the mean and std over all the images.

Requires:
 * OpenCV
 * tqdm

In this example, I use `Stanford Dogs Datasets" (~20k images)

Example Outputs: (8 cores CPU i7 4970K)
(env) dat@desktop:****/StanfordDogs$ python compute_mean_std.py 
100%|██████████████████████████████████████████████████████████| 20580/20580 [01:06<00:00, 308.06it/s]
[0.4761392  0.45182742 0.39101657] [0.23364353 0.2289059  0.22732813]
""""
import os
import cv2
import tqdm
import multiprocessing as mp

import scipy.io as sio   # read matlab file
import numpy as np

def compute_func(filename):
  img = cv2.cvtColor(cv2.imread(filename), cv2.COLOR_BGR2RGB) / 255.
  return np.mean(img, (0, 1)), np.std(img, (0, 1))

def main():

  # Read filenames
  matfile = sio.loadmat('file_list.mat')
  fnames = [os.path.join('Images', str(i[0][0])) for i in matfile['file_list']]

  with mp.Pool(mp.cpu_count()) as pool:
    result = list(tqdm.tqdm(pool.imap(
        func=compute_func,
        iterable=[i for i in fnames]), total=len(fnames)))
    mean,  std = zip(*result)

  print(np.mean(mean, 0), np.mean(std, 0))

if __name__ == '__main__':
  main()