Array conversion using scikit-image: from integer

问题:

I am facing some kind of problem when converting an integer image to a float image using scikit-image.

This is an example (the image is a 2 pixel image):

from numpy import array,uint8;
import skimage;

rgb = array([array([[0,0,0],[0,0,5]])]) 
i1 = skimage.img_as_float(rgb)#rgb.dtype ->dtype('int32')
i2 = skimage.img_as_float(rgb.astype(uint8))
i3 = skimage.img_as_float(rgb.astype(float))

print i1[0,1,:]
print i2[0,1,:]
print i3[0,1,:]

I expected this:

[ 0.  0.  5.]
[ 0.  0.  5.]
[ 0.  0.  5.]

But I got this:

[  2.32830644e-10   2.32830644e-10   2.56113708e-09]
[ 0.          0.          0.01960784]
[ 0.  0.  5.]

It is normal to loss precision from float to int, but here I am losing the real information when passing from int to float using img_as_float. I didn't found anything when reading the code on GitHub...

Why is this possible?

回答1:

img_as_float() is not just type conversion, it convert full unsigned integer range to [0, 1], full signed integer range to [-1, 1].

• i1, the dtype is int32, means convert [-2147483648, 2147483647] to [-1, 1]
• i2, the dtype is uint8, means convert [0, 255] to [0, 1]
• i3, because the dtype is already float, do nothing.