NumPy Image Data Conventions

Image Data Axes

NumPy and other Python based frameworks - like TensorFlow - define image data as arrays of pixels.

This means, that an image is represented by a 3-dimensional array with:

• axis 0 representing the image height,
• axis 1 representing the image width,
• axis 2 representing the pixel depth (number of color channels)

Following this convention, the Python data shape of an image looks like this:

• (height, width, depth)

References

• Image axes in TensorFlow
• Unfortunately, PyTorch defines image axes in a different way
• Python Pillow - Image Module

Reading an Image File into a NumPy Array

from PIL import Image
import numpy as np

img = Image.open("test.png")
data = np.array(img)

Reading all Image Files from a Zip Package

from zipfile import ZipFile
from PIL import Image
import numpy as np

zf = ZipFile("test.zip")

data = []
for zfi in zf.infolist():
    fi = zf.open(zfi)
    img = Image.open(fi)
    data.append(np.array(img))