pltrdy pltrdy - 11 months ago 86
Python Question

Grayscale convertion using tf.reduce_mean & tf.concat

TensorFlow newbie here, training on a simple tutorial which I just fail.
The point is to convert an image to grayscale.

Our data is basically an

(height of the picture, width, and color on three values r,g,b).

So it might be equivalent to transform each array cell from
[r, g, b]
[gray, gray, gray]
gray = mean(r, g, b)

Thus I checked the doc for a mean function and found reduce_mean.
I used it on the color axis, i.e. axis=2, then concat the result on itself using axis 2 again to "replicate" mean value and finally get 3 times the gray value (=mean) as red, green and blue.

See the code below:

import tensorflow as tf
import matplotlib.image as mpimg

filename = "MarshOrchid.jpg"
raw_image_data = mpimg.imread(filename)

image = tf.placeholder("uint8", [None, None, 3])

# Reduce axis 2 by mean (= color)
# i.e. image = [[[r,g,b], ...]]
# out = [[[ grayvalue ], ... ]] where grayvalue = mean(r, g, b)
out = tf.reduce_mean(image, 2, keep_dims=True)

# Associate r,g,b to the same mean value = concat mean on axis 2.
# out = [[[ grayvalu, grayvalue, grayvalue], ...]]
out = tf.concat(2, [out, out, out])

with tf.Session() as session:
result =, feed_dict={image: raw_image_data})


(You can get the image here)

This code can be executed but the result isn't ok.

When grayscale fails

Wondering what happened I check my variables, and it turns out that the mean isn't ok, has shown on screenshot below, mean(147, 137, 88) != 38

enter image description here

Any ideas?
Can't figure out what I did wrong...


Answer Source

The error come from the dtype of your placeholder. Cause the type inference, intermediate tensors cannot have values greater than 255 (2^8-1). When Tensorflow compute mean(147, 137, 88), first it compute : sum(147, 137, 88)=372, but 372>256 so it keep 372% 256 = 116.

And so mean(147, 137, 88) = sum(147, 137, 88)/3 = 116/3 = 40. Change the dtype of your placeholder to "uint16" or "uint32".