note.nkmk.me

Image processing with Python, NumPy

Posted: 2019-05-14 / Modified: 2020-10-20 / Tags: Python, NumPy, Image Processing

By reading the image as a NumPy array ndarray, various image processing can be performed using NumPy functions.

By the operation of ndarray, you can get and set (change) pixel values, trim images, concatenate images, etc. Those who are familiar with NumPy can do various image processing without using libraries such as OpenCV.

Even when using OpenCV, OpenCV for Python treats image data as ndarray, so it is useful to know how to use NumPy (ndarray). In addition to OpenCV, there are many libraries such as scikit-image that treat images as ndarray.

This article describes the following contents.

Read and write images:

  • How to read image file as NumPy array ndarray
  • How to save NumPy array ndarray as image file

Examples of image processing with NumPy (ndarray):

  • Get and set (change) pixel values
  • Generation of single color image and concatenation
  • Negative / positive inversion (inversion of pixel value)
  • Color reduction
  • Binarization
  • Gamma correction
  • Trimming with slice
  • Split with slice or function
  • Paste with slice
  • Alpha blending and masking
  • Rotate and flip

Sample codes on this article use Pillow to read and save image files. If you want to use OpenCV, see the following article.

See also the following article about Pillow. Simple operations such as reading and saving images, resizing and rotating images can be done by Pillow alone.

Sponsored Link

How to read an image file as ndarray

Take the following image as an example.

lena

Passing the image data read by PIL.Image.open() to np.array() returns 3D ndarray whose shape is (row (height), column (width), color (channel)).

from PIL import Image
import numpy as np

im = np.array(Image.open('data/src/lena.jpg'))

print(type(im))
# <class 'numpy.ndarray'>

print(im.dtype)
# uint8

print(im.shape)
# (225, 400, 3)

The order of colors (channels) is RGB (red, green, blue). Note that it is different from the case of reading with cv2.imread() of OpenCV.

If you convert the image to grayscale with convert('L') and then pass it to np.array(), it returns 2D ndarray whose shape is (row (height), column (width)).

im_gray = np.array(Image.open('data/src/lena.jpg').convert('L'))

print(im_gray.shape)
# (225, 400)

You can also get ndarray from PIL.Image with np.asarray(). np.array() returns a rewritable ndarray, while np.asarray() returns a non-rewritablendarray.

For np.array(), you can change the value of the element (pixel).

print(im.flags.writeable)
# True

print(im[0, 0, 0])
# 109

im[0, 0, 0] = 0

print(im[0, 0, 0])
# 0

For np.asarray(), you cannot change value because rewriting is prohibited. It is possible to create a new ndarray based on the read ndarray.

im_as = np.asarray(Image.open('data/src/lena.jpg'))

print(type(im_as))
# <class 'numpy.ndarray'>

print(im_as.flags.writeable)
# False

# im_as[0, 0, 0] = 0
# ValueError: assignment destination is read-only

The data type dtype of the read ndarray is uint8 (8-bit unsigned integer).

If you want to process it as a floating point number float, you can convert it with astype() or specify the data type in the second argument of np.array() and np.asarray().

im_f = im.astype(np.float64)
print(im_f.dtype)
# float64

im_f = np.array(Image.open('data/src/lena.jpg'), np.float64)
print(im_f.dtype)
# float64

How to save NumPy array ndarray as image file

Passing ndarray to Image.fromarray() returns PIL.Image. It can be saved as an image file with save() method. The format of the saved file is automatically determined from the extension of the path passed in the argument of save().

pil_img = Image.fromarray(im)
print(pil_img.mode)
# RGB

pil_img.save('data/temp/lena_save_pillow.jpg')

A grayscale image (2D array) can also be passed to Image.fromarray(). mode automatically becomes 'L' (grayscale). It can be saved with save().

pil_img_gray = Image.fromarray(im_gray)
print(pil_img_gray.mode)
# L

pil_img_gray.save('data/temp/lena_save_pillow_gray.jpg')

If you just want to save it, you can write it in one line.

Image.fromarray(im).save('data/temp/lena_save_pillow.jpg')
Image.fromarray(im_gray).save('data/temp/lena_save_pillow_gray.jpg')

If the data type dtype of ndarray is float, etc., an error will occur, so it is necessary to convert to uint8.

# pil_img = Image.fromarray(im_f)
# TypeError: Cannot handle this data type

pil_img = Image.fromarray(im_f.astype(np.uint8))
pil_img.save('data/temp/lena_save_pillow.jpg')

Note that if the pixel value is represented by 0.0 to 1.0, it is necessary to multiply by 255 and convert to uint8 and save.

With save(), parameters according to the format can be passed as arguments. See Image file format for details.

For example, in the case of JPG, you can pass the quality of the image to the argument quality. It ranges from 1 (the lowest) to 95 (the highest) and defaults to 75.

Get and set (change) pixel values

You can get the value of a pixel by specifying the coordinates at the index [row, columns] of ndarray. Note that the order is y, x in xy coordinates. The origin is the upper left.

from PIL import Image
import numpy as np

im = np.array(Image.open('data/src/lena.jpg'))

print(im.shape)
# (225, 400, 3)

print(im[100, 150])
# [111  81 109]

print(type(im[100, 150]))
# <class 'numpy.ndarray'>

The above example shows the value at (y, x) = (100, 150), i.e. the 100th row and 150th column of pixels. As mentioned above, the colors of the ndarray obtained using Pillow are in RGB order, so the result is (R, G, B) = (111, 81, 109).

You can also use unpack to assign them to separate variables.

R, G, B = im[100, 150]

print(R)
# 111

print(G)
# 81

print(B)
# 109

It is also possible to get the value by specifying the color.

print(im[100, 150, 0])
# 111

print(im[100, 150, 1])
# 81

print(im[100, 150, 2])
# 109

You can also change to a new value. You can change RGB all at once, or you can change it with just a single color.

im[100, 150] = (0, 50, 100)

print(im[100, 150])
# [  0  50 100]

im[100, 150, 0] = 150

print(im[100, 150])
# [150  50 100]

Generation of single color image and concatenation

Generate single-color images by setting other color values to 0, and concatenate them horizontally with np.concatenate(). You can also concatenate images using np.hstack() or np.c_[]

from PIL import Image
import numpy as np

im = np.array(Image.open('data/src/lena_square.png'))

im_R = im.copy()
im_R[:, :, (1, 2)] = 0
im_G = im.copy()
im_G[:, :, (0, 2)] = 0
im_B = im.copy()
im_B[:, :, (0, 1)] = 0

im_RGB = np.concatenate((im_R, im_G, im_B), axis=1)
# im_RGB = np.hstack((im_R, im_G, im_B))
# im_RGB = np.c_['1', im_R, im_G, im_B]

pil_img = Image.fromarray(im_RGB)
pil_img.save('data/dst/lena_numpy_split_color.jpg')
NumPy image processing split color

Negative / positive inversion (invert pixel value)

It is also easy to calculate and manipulate pixel values.

A negative-positive inverted image can be generated by subtracting the pixel value from the max value (255 for uint8).

import numpy as np
from PIL import Image

im = np.array(Image.open('data/src/lena_square.png').resize((256, 256)))

im_i = 255 - im

Image.fromarray(im_i).save('data/dst/lena_numpy_inverse.jpg')

Python NumPy inverse

Because the original size is too large, it is resized with resize() for convenience. The same applies to the following examples.

Color reduction

Cut off the remainder of the division using // and multiply again, the pixel values become discrete and the number of colors can be reduced.

import numpy as np
from PIL import Image

im = np.array(Image.open('data/src/lena_square.png').resize((256, 256)))

im_32 = im // 32 * 32
im_128 = im // 128 * 128

im_dec = np.concatenate((im, im_32, im_128), axis=1)

Image.fromarray(im_dec).save('data/dst/lena_numpy_dec_color.png')

Python NumPy decrease color

Binarization

It is also possible to assign to black and white according to the threshold.

See the following articles for details.

Python NumPy OpenCV binarization

Sponsored Link

Gamma correction

You can do anything you want with pixel values, such as multiplication, division, exponentiation, etc.

You don't need to use the for loop because the entire image can be calculated as it is.

from PIL import Image
import numpy as np

im = np.array(Image.open('data/src/lena_square.png'))

im_1_22 = 255.0 * (im / 255.0)**(1 / 2.2)
im_22 = 255.0 * (im / 255.0)**2.2

im_gamma = np.concatenate((im_1_22, im, im_22), axis=1)

pil_img = Image.fromarray(np.uint8(im_gamma))
pil_img.save('data/dst/lena_numpy_gamma.jpg')
NumPy image processing split gamma

As a result of the calculation, the data type dtype of numpy.ndarray is converted to the floating point number float. Note that you need to convert it to uint8 when you save it.

Trimming with slice

By specifying an area with slice, you can trim it to a rectangle.

from PIL import Image
import numpy as np

im = np.array(Image.open('data/src/lena_square.png'))

print(im.shape)
# (512, 512, 3)

im_trim1 = im[128:384, 128:384]
print(im_trim1.shape)
# (256, 256, 3)

Image.fromarray(im_trim1).save('data/dst/lena_numpy_trim.jpg')

numpy image trimming 1

See the following article for more information on slicing for numpy.ndarray.

It may be convenient to define a function that specifies the upper left coordinates and the width and height of the area to be trimmed.

def trim(array, x, y, width, height):
    return array[y:y + height, x:x+width]

im_trim2 = trim(im, 128, 192, 256, 128)
print(im_trim2.shape)
# (128, 256, 3)

Image.fromarray(im_trim2).save('data/dst/lena_numpy_trim2.jpg')

numpy image trimming 2

If you specify outside the size of the image, it will be ignored.

im_trim3 = trim(im, 128, 192, 512, 128)
print(im_trim3.shape)
# (128, 384, 3)

Image.fromarray(im_trim3).save('data/dst/lena_numpy_trim3.jpg')

numpy image trimming 3

Split with slice or function

You can also split the image by slicing.

from PIL import Image
import numpy as np

im = np.array(Image.open('data/src/lena_square.png').resize((256, 256)))

print(im.shape)
# (256, 256, 3)

im_0 = im[:, :100]
im_1 = im[:, 100:]

print(im_0.shape)
# (256, 100, 3)

print(im_1.shape)
# (256, 156, 3)

Image.fromarray(im_0).save('data/dst/lena_numpy_split_0.jpg')
Image.fromarray(im_1).save('data/dst/lena_numpy_split_1.jpg')

numpy image split 0

numpy image split 1

It is also possible to split the image with NumPy function.

np.hsplit() splits ndarray horizontally. If an integer value is specified for the second argument, ndarray is splitted equally.

im_0, im_1 = np.hsplit(im, 2)

print(im_0.shape)
# (256, 128, 3)

print(im_1.shape)
# (256, 128, 3)

If a list is specified as the second argument, ndarray is splitted at the position of that values.

im_0, im_1, im_2 = np.hsplit(im, [100, 150])

print(im_0.shape)
# (256, 100, 3)

print(im_1.shape)
# (256, 50, 3)

print(im_2.shape)
# (256, 106, 3)

np.vsplit() splits ndarray vertically. The usage of np.vsplit() is the same as np.hsplit().

When an integer value is specified as the second argument with np.hsplit() or np.vsplit(), an error will occur if it cannot be splitted equally. np.array_split() adjusts the size appropriately and splits it.

# im_0, im_1, im_2 = np.hsplit(im, 3)
# ValueError: array split does not result in an equal division

im_0, im_1, im_2 = np.array_split(im, 3, axis=1)

print(im_0.shape)
# (256, 86, 3)

print(im_1.shape)
# (256, 85, 3)

print(im_2.shape)
# (256, 85, 3)

Paste with slice

Using slices, one array rectangle can be replaced with another array rectangle.

By using this, a part of the image or the entire image can be pasted to another image.

import numpy as np
from PIL import Image

src = np.array(Image.open('data/src/lena_square.png').resize((128, 128)))
dst = np.array(Image.open('data/src/lena_square.png').resize((256, 256))) // 4

dst_copy = dst.copy()
dst_copy[64:128, 128:192] = src[32:96, 32:96]

Image.fromarray(dst_copy).save('data/dst/lena_numpy_paste.jpg')

numpy image paste

dst_copy = dst.copy()
dst_copy[64:192, 64:192] = src

Image.fromarray(dst_copy).save('data/dst/lena_numpy_paste_all.jpg')

numpy image paste all

Note that an error will occur if the size of the area specified on the left side differs from the size of the area specified on the right side.

Alpha blending and masking

By the operation for each element (= pixel) of the array, two images can be alpha-blended or composited based on a mask image. See the following articles for details.

NumPy image alpha blend gradation

NumPy image blend blur

Rotate and flip

There are also functions that rotate the array and flip it up, down, left and right.

Original image:

lena

Rotated image:

numpy rot90 image

Flipped image:

nupmy flipud image

Sponsored Link
Share

Related Categories

Related Articles