Importing PNG files into Numpy?

Question

I have about 200 grayscale PNG images stored within a directory like this.

1.png
2.png
3.png
...
...
200.png

I want to import all the PNG images as NumPy arrays. How can I do this?

Solution 1

Using just scipy, glob and having PIL installed (pip install pillow) you can use scipy's imread method:

from scipy import misc
import glob

for image_path in glob.glob("/home/adam/*.png"):
    image = misc.imread(image_path)
    print image.shape
    print image.dtype

UPDATE

According to the doc, scipy.misc.imread is deprecated starting SciPy 1.0.0, and will be removed in 1.2.0. Consider using imageio.imread instead. See the answer by Charles.

Solution 2

According to the doc, scipy.misc.imread is deprecated starting SciPy 1.0.0, and will be removed in 1.2.0. Consider using imageio.imread instead.

Example:

import imageio

im = imageio.imread('my_image.png')
print(im.shape)

You can also use imageio to load from fancy sources:

im = imageio.imread('http://upload.wikimedia.org/wikipedia/commons/d/de/Wikipedia_Logo_1.0.png')

Edit:

To load all of the *.png files in a specific folder, you could use the glob package:

import imageio
import glob

for im_path in glob.glob("path/to/folder/*.png"):
     im = imageio.imread(im_path)
     print(im.shape)
     # do whatever with the image here

Solution 3

This can also be done with the Image class of the PIL library:

from PIL import Image
import numpy as np

im_frame = Image.open(path_to_file + 'file.png')
np_frame = np.array(im_frame.getdata())

Note: The .getdata() might not be needed - np.array(im_frame) should also work

Solution 4

Using a (very) commonly used package is prefered:

import matplotlib.pyplot as plt
im = plt.imread('image.png')

Solution 5

If you are loading images, you are likely going to be working with one or both of matplotlib and opencv to manipulate and view the images.

For this reason, I tend to use their image readers and append those to lists, from which I make a NumPy array.

import os
import matplotlib.pyplot as plt
import cv2
import numpy as np

# Get the file paths
im_files = os.listdir('path/to/files/')

# imagine we only want to load PNG files (or JPEG or whatever...)
EXTENSION = '.png'

# Load using matplotlib
images_plt = [plt.imread(f) for f in im_files if f.endswith(EXTENSION)]
# convert your lists into a numpy array of size (N, H, W, C)
images = np.array(images_plt)

# Load using opencv
images_cv = [cv2.imread(f) for f in im_files if f.endswith(EXTENSION)]
# convert your lists into a numpy array of size (N, C, H, W)
images = np.array(images_cv)

The only difference to be aware of is the following:

• opencv loads channels first
• matplotlib loads channels last.

So a single image that is 256*256 in size would produce matrices of size (3, 256, 256) with opencv and (256, 256, 3) using matplotlib.

View more solutions

Comments

[pbu]

Just change to glob.glob("./train/*.png")

[SherylHohman]

Perfect. Great All-In-One Solution. I had stored images into an np.array, but then ran into trouble as the array had (shape == (num_images,) with each image (shape == (32,32,3)). Your solution (plus im = np.reshape(num_images,32,32,3) works brilliantly! :-)

[SherylHohman]

typo: I don't even need the reshape call above. In mine vexed hack, massaging it into that desired shape was getting messy. Thanks for the direct path.

[Charles]

Downvoter, if you could please help me make this answer better by telling me what to improve, it would be very appreciated!

[devssh]

I'm not downvoter but the question is to load a list of images in a folder. Can you modify your answer to reflect that, not just 1 - 200.png but what if the png have random names, that would help me a lot. I'm sure I can use os to ls and get the file names but is there a better way? Perhaps you should edit to add glob

[devssh]

Also, remember to add a try catch as imread may throw ValueError. I don't have privileges to edit or I would have updated it for you. :)

[Charles]

The developer should choose if, in this context, the exception should be raised and halt execution or handled in a specific way. Without context, raising is preferred.

[Charles]

well, glob only finds files in your directory tree that match a certain "globbing" path (one that can contains wildcards etc.) so I'm not sure what you mean by "invalid png errors are ignored by glob" since glob only looks at filenames and know nothing about your images.

[devssh]

Yes, it should, but glob saved me the trouble by not reading invalid png for reasons unknown. I used imageio.imwrite along with skimage.transform.resize and it fits into numpy!

[devssh]

I found out what was happening, my try except caught the error when the glob on invalid png format blew up.

[EL_DON]

I already have matplotlib imported most of the time, so this has basically no extra cost. Did matplotlib not always have imread or something? What's the advantage of the packages used in other answers?

[Rob]

I don't think the .getdata() is even needed. np.array(im_frame) should also work.

[mrk]

Thanks for the contribution: I added your remark as a comment to my answer for highlighting.

[sailfish009]

i had vote up, but .getdata() cause error, so had to use without .getdata() (Rob's method) both code is not same! : raise TypeError('Input type {} is not supported'.format(npimg.dtype)) TypeError: Input type int64 is not supported

[mrk]

seems as if you were having an issue with your input type.

[liang]

the doc link is broken

[Charles]

@liang updated the link; thanks for pointing it out

[Brandon]

To get it as RGB use img = np.asarray(Image.open('file.png').convert('RGB'))

[milembar]

no advantage. Just a convenience wrapper over pillow (PIL) to sparse one extra line of code.

[Nir]

@milembar as EL_DON mentioned many scripts already have matplotlib imported

[C-3PO]

I ran a few benchmarks, and it seems that np.asarray(PIL.Image.open(file.png)) is slightly faster than imageio.imread(...).