[PYTHON] PIL / Pillow cheat sheet

PIL / Pillow is a compact and fast image library for Python. We have summarized frequently used processes (updated from time to time)

Difference between PIL and Pillow

There is basically no reason to use PIL, Pillow has a bugfix for resizing filters and is of higher quality.

About Pillow speed

Pillow is tuned very fast and always runs faster than its similar library, ImageMagick. However, getpixel / putpixel is very slow, so don't use it for anything other than image generation. There is also a faster pillow-simd. It seems to be about 4 to 5 times faster than the original Pillow.

pillow-simd https://github.com/uploadcare/pillow-simd

reference

https://python-pillow.org/pillow-perf/

Image mode list

Resize filter

• Image.ANTIALIAS is another name for Image.LANCZOS and is left for compatibility. reference http://pillow.readthedocs.io/en/3.4.x/handbook/concepts.html#concept-filters

Image module

Grayscale

img.convert("L")

Grayscale considering alpha value

alpha.convert("LA")

→

By the way, convert ('L') does not consider the alpha value.

alpha.convert("L")

→

HSV conversion

img.convert("HSV")

Only pillow can be converted to HSV color space. It is composed of three components: Hue, Saturation, and Value. The following is an example of shifting the color wheel.

h, s, v = img.convert("HSV").split()
_h = ImageMath.eval("(h + 128) % 255", h=h).convert("L")
Image.merge("HSV", (_h, s, v)).convert("RGB")

CIE XYZ conversion

CIE XYZ is a color space adjusted so that the Euclidean distance between colors is the same as the difference perceived by humans.

rgb2xyz = (
    0.412453, 0.357580, 0.180423, 0,
    0.212671, 0.715160, 0.072169, 0,
    0.019334, 0.119193, 0.950227, 0
)
img.convert("RGB", rgb2xyz)

Binarization

gray = img.convert("L")                     #Convert to grayscale
gray.point(lambda x: 0 if x < 230 else x)   #If the value is 230 or less, it will be 0.

Brighten / darken the image

img.point(lambda x: x * 1.5)    # 1.Make it 5 times brighter
img.point(lambda x: x * 0.5)    # 1 /Darken to 2

Sepia

Convert the image to grayscale and then sepia it.

gray = img.convert("L")
Image.merge(
    "RGB",
    (   
        gray.point(lambda x: x * 240 / 255),
        gray.point(lambda x: x * 200 / 255),
        gray.point(lambda x: x * 145 / 255)
    )
)

Gamma correction

Gamma correction can also be converted at a very high speed by using a look-up table. Assuming that src = input color, γ = gamma value, and g = gain value, the gamma correction formula is as follows.

dst = \biggl(\frac{src}{255}\biggr)^{1/γ} \times g \times 255

def gamma_table(gamma_r, gamma_g, gamma_b, gain_r=1.0, gain_g=1.0, gain_b=1.0):
    r_tbl = [min(255, int((x / 255.) ** (1. / gamma_r) * gain_r * 255.)) for x in range(256)]
    g_tbl = [min(255, int((x / 255.) ** (1. / gamma_g) * gain_g * 255.)) for x in range(256)]
    b_tbl = [min(255, int((x / 255.) ** (1. / gamma_b) * gain_b * 255.)) for x in range(256)]
    return r_tbl + g_tbl + b_tbl

img.point(gamma_table(1.2, 0.5, 0.5))

Speeding up point used in loops

It is better not to pass lambda to ʻImage.point in the loop, it is faster to expand it in advance because the argument passed to point` is just a conversion table.

for ...:
   img.point(lambda x: x * 100)

#The lower process is equal to the upper process but faster
table = [x * 100 for x in range(256)] * len(img.getbands())
for ...:
    img.point(table)

getbbox ʻImage.getbboxreturns the smallest non-zero value in the image, an image with all zeros returnsNone`.

Margin cut of alpha component

alpha = Image.open("alpha.png ")

crop = alpha.split()[-1].getbbox()
alpha.crop(crop)

→

Same pixel check

If the two images are the same ʻImageChops.difference returns all 0 images, so if getbboxisNone`, it can be judged to be the same.

ImageChops.difference(img1, img2).getbbox() is None

resize

img.resize((128, 128), Image.LANCZOS)

thumbnail

Thumbnails maintain aspect ratio, unlike resizing. Note that for some reason thumbnail is a destructive method, it's ʻImage.copy` so it's a good idea to make a duplicate.

img.thumbnail((128, 128), Image.LANCZOS)
img.size
# (128, 79)

rotation

Specifying True for the argument ʻexpand` expands the image if it grows when rotated.

img.rotate(90, expand=True)

Mosaic processing

Mosaic processing can be reduced and enlarged with ʻImage.LINEAR`, but if you reduce it after applying Gaussian blur, it will be a soft mosaic.

#Jagged mosaic
img.resize([x // 8 for x in img.size]).resize(img.size)

#Apply Gaussian blur for a soft mosaic
gimg = img.filter(ImageFilter.GaussianBlur(4))
gimg.resize([x // 8 for x in img.size]).resize(img.size)

Alpha blend

Image.blend(img,effect_img, 0.5)

Subtractive color

img.quantize(4)    #Color reduction to 4 colors

Paste the alpha image

To paste an image with alpha, specify the image with alpha in the argument'mask'of ʻImage.paste`.

img.paste(alpha, mask=alpha)

Count the colors used

ʻImage.getcolors`, which counts the colors used, cannot count more than 255 colors with no arguments. For images that use more than 255 colors, it is safe to pass the number of pixels as an argument. [^ 1]

[^ 1]: 2017-02-07 Fixed It was'Image.getcount', but it is correctly'Image.getcolors', I will fix it.

img.getcolors(img.size[0] * img.size[1])

histogram

Returns a list of image color histograms. Since each band is returned in succession, 256 x 3 = 768 elements are returned in RGB mode.

img.histogram()

Color replacement

There is no method to replace the color, if you want to replace the color please refer to the article below.

Replace image colors fast with PIL / Pillow

ImageOps module

Negative / positive reversal

ImageOps.invert(img)

Flip left / right / flip up / down

ImageOps.mirror(img)    #Flip horizontal
ImageOps.flip(img)      #flip upside down

Colorization

Colors a grayscale image with a pixel value of 0 to black and a pixel value of 255 to white.

gray = ImageOps.grayscale(img)
ImageOps.colorize(gray, black=(0, 0, 0), white=(255, 255, 0))

　→　

Posterize

Reduces the bit depth of the image to the value of the argument to simplify the color.

ImageOps.posterize(img, 2)

Solarize

Inverts all pixel values above the threshold. I don't know where to use it.

ImageOps.solarize(img, 128)

Equalize

Equalize the histogram of the image. Apply nonlinear mapping to the input image to create a uniform distribution of grayscale values in the output image.

ImageOps.equalize(img)

ImageChops module

The ʻImageChops` module is a module for manipulating channels.

The left is the image to be affected and the right is the image for effects. In this chapter, we will use these two images as samples.

Dodge (Linear) / Subtraction

ImageChops.add(img, effect_img)         # img + effect_img
ImageChops.subtract(img, effect_img)    # img - effect_img

mod operation

ImageChops.add_modulo(img, effect_img)         # img + effect_img % MAX
ImageChops.subtract_modulo(img, effect_img)    # img - effect_img % MAX

Multiply / screen

ImageChops.multiply(img, effect_img)
ImageChops.screen(img, effect_img)

Comparison (bright) / comparison (dark)

ImageChops.lighter(img, effect_img)
ImageChops.darker(img, effect_img)

Absolute value of difference

ImageChops.difference(img, effect_img)

offset

ImageChops.offset(img, 100, 100)

ImageFilter module

Performs a convolution (convolution operation). Various image conversions are performed by rearranging the matrix called the kernel.

reference

https://github.com/python-pillow/Pillow/blob/6e7553fb0f12025306b2819b9b842adf6b598b2e/PIL/ImageFilter.py

ImageFilter.BLUR

img.filter(ImageFilter.BLUR)

# size: (5, 5),
# scale: 16,
# offset: 0,
# kernel:(
#     1,  1,  1,  1,  1,
#     1,  0,  0,  0,  1,
#     1,  0,  0,  0,  1,
#     1,  0,  0,  0,  1,
#     1,  1,  1,  1,  1
# )

ImageFilter.DETAIL

img.filter(ImageFilter.DETAIL)

# size: (3, 3), 
# scale: 6,
# offset: 0,
# kernel: (
#     0, -1,  0,
#     -1, 10, -1,
#     0, -1,  0
# )

ImageFilter.SHAPEN

img.filter(ImageFilter.SHARPEN)

# size: (3, 3),
# scale: 16,
# offset: 0,
# kernel: (
#     -2, -2, -2,
#     -2, 32, -2,
#     -2, -2, -2
# )

ImageFilter.CONTOUR

img.filter(ImageFilter.CONTOUR)

# size: (3, 3),
# scale: 1,
# offset: 255,
# kernel: (
#     -1, -1, -1,
#     -1,  8, -1,
#     -1, -1, -1
# )

ImageFilter.EDGE_ENHANCE / ImageFilter.EDGE_ENHANCE_MORE

img.filter(ImageFilter.EDGE_ENHANCE)

# size: (3, 3),
# scale: 2,
# offset: 0,
# kernel: (
#     -1, -1, -1,
#     -1, 10, -1,
#     -1, -1, -1
# )

img.filter(ImageFilter.EDGE_ENHANCE_MORE)

# size: (3, 3),
# scale: 1,
# offset: 0,
# kernel: (
#     -1, -1, -1,
#     -1,  9, -1,
#     -1, -1, -1
# )

ImageFilter.EMBOSS

img.filter(ImageFilter.EMBOSS)

# size: (3, 3),
# scale: 1,
# offset: 128,
# kernel: (
#     -1,  0,  0,
#     0,  1,  0,
#     0,  0,  0
# )

ImageFilter.FIND_EDGES

img.filter(ImageFilter.FIND_EDGES)

# size: (3, 3),
# scale: 1,
# offset: 0,
# kernel: (
#     -1, -1, -1,
#     -1,  8, -1,
#     -1, -1, -1
# )

ImageFilter.SMOOTH / ImageFilter.SMOOTH_MORE

img.filter(ImageFilter.SMOOTH)
# size: (3, 3),
# scale: 13,
# offset: 0,
# kernel: (
#     1,  1,  1,
#     1,  5,  1,
#     1,  1,  1
# )
# 

img.filter(ImageFilter.SMOOTH_MORE)
# size: (5, 5),
# scale: 100,
# offset: 0,
# kernel: (
#     1,  1,  1,  1,  1,
#     1,  5,  5,  5,  1,
#     1,  5, 44,  5,  1,
#     1,  5,  5,  5,  1,
#     1,  1,  1,  1,  1
# )

Gaussian blur

Gaussian Blurにより画面の平滑化します。

img.filter(ImageFilter.GaussianBlur(1.0))
img.filter(ImageFilter.GaussianBlur(1.5))
img.filter(ImageFilter.GaussianBlur(3.0))

Expansion / contraction

MaxFilter is called Dilation and MinFilter is called Erosion.

img.filter(ImageFilter.MinFilter())
img.filter(ImageFilter.MaxFilter())

reference

Expansion / contraction / opening / closing

Median filter

MedianFilter is often used to remove noise, and its contours are less blurred than Gaussian filters.

img.filter(ImageFilter.MedianFilter())

→

reference

Noise reduction

Mode filter

Selects the most frequently used pixel value in a box of the specified size. Pixel values that occur only once or twice are ignored. ~~ I don't know where to use it. See ~~ Make your photos pictorial with Pillow's Mode Filter.

img.filter(ImageFilter.ModeFilter(5))

Image Enhance module

Color balance adjustment

enhancer = ImageEnhance.Color(img)
enhancer.enhance(0.0)    #Black and white
enhancer.enhance(0.5)    #  ↕
enhancer.enhance(1.0)    #The original image

Contrast adjustment

enhancer = ImageEnhance.Contrast(img)
enhancer.enhance(0.0)    #Gray image
enhancer.enhance(0.5)    #  ↕
enhancer.enhance(1.0)    #The original image

Brightness adjustment

enhancer = ImageEnhance.Brightness(img)
enhancer.enhance(0.0)    #Black image
enhancer.enhance(0.5)    #  ↕
enhancer.enhance(1.0)    #The original image

Sharpness adjustment

enhancer = ImageEnhance.Sharpness(img)
enhancer.enhance(0.0)    #Blurred image
enhancer.enhance(0.5)    #  ↕
enhancer.enhance(1.0)    #The original image
enhancer.enhance(1.5)    #  ↕
enhancer.enhance(2.0)    #Sharp image

ImageMath module

The ʻImageMath` module is a module that allows you to write operations between pixels as if they were numerical operations. If you master it, you will be able to easily write complicated image processing.

--Only single band can be calculated, multi band is processed after splitting with Image.split --The range of values when calculating with float is 0.0 to 255.0 instead of 0.0 to 1.0. --The mode of the image being calculated will be "I" (int) or "F" (float), and finally the mode will be converted to "L". --Various operations (+,-, \ *, /, * \ *,%) are not pixel processing but whole image operation --The operation is performed for each Image, not for each pixel, and an Image is generated for each operation. --You can also pass callable objects such as lambda instead of operations

Since it can only process single bands, it is troublesome when converting images such as RGB, so it is a good idea to prepare the following helper function.

def _blend_f(img1, img2, func):
    blend_eval = "convert(func(float(a), float(b)), 'L')"
    bands = [
        ImageMath.eval(
            blend_eval,
            a=a,
            b=b,
            func=func
        )
        for a, b in zip(img1.split(), img2.split())
    ]
    return Image.merge(img1.mode, bands)

reference

Implement drawing modes such as PhotoShop at high speed with PIL / Pillow

overlay

def _over_lay(a, b):
    _cl = 2 * a * b / 255
    _ch = 2 * (a + b - a * b / 255) - 255
    return _cl * (a < 128) + _ch * (a >= 128)

_blend_f(img, effect_img, _over_lay)

Soft light

def _soft_light(a, b):
    _cl = (a / 255) ** ((255 - b) / 128) * 255
    _ch = (a / 255) ** (128 / b) * 255
    return _cl * (b < 128) + _ch * (b >= 128)

_blend_f(img, effect_img, _soft_light)

Hard light

def _hard_light(a, b):
    _cl = 2 * a * b / 255
    _ch = 2.0 * (a + b - a * b / 255.0) - 255.0
    return _cl * (b < 128) + _ch * (b >= 128)

_blend_f(img, effect_img, _hard_light)

Use Photoshop drawing mode

I am creating a module called "Image4Layer" that implements the drawing mode of Photoshop.

https://github.com/pashango2/Image4Layer

Installation is easy with pip, pillow (PIL) must be pre-installed to run.

$pip install image4layer

It's easy to use, it's an example of compositing in color-dodge mode.

from PIL import Image
from image4layer import Image4Layer

source = Image.open("ducky.png ")
backdrop = Image.open("backdrop.png ")

Image4Layer.color_dodge(backdrop, source)

GIF writing

You can write multiple GIFs (GIF animations).

im.save(out, save_all=True, append_images=[im1, im2, ...])

This is an example of creating a simple animated GIF.

imgs = []
for i in range(100):
    imgs.append(img.point(lambda x: x * (1.0 - (i/100))))
    
img.save("anime.gif", save_all=True, append_images=imgs, loop=True)

reference

http://pillow.readthedocs.io/en/4.0.x/handbook/image-file-formats.html?highlight=seek#saving

Conversion to QImagae

Converting PyQt to QImage uses the ʻImageQt` module.

ImageQt.ImageQt(img)

If you are using PySide, the following method is recommended.

from PySide.QtGui import *
import io

img_buffer = io.BytesIO()
base.save(img_buffer, "BMP")
qimage = QImage()
qimage.loadFromData(img_buffer.getvalue(), "BMP")

It may seem like a wasteful process, but Pillow / PySide will take care of the troublesome parts such as RGB → BGR conversion and Y-axis inversion problem.

reference

Mutual conversion between PIL.Image and PyQt4.QtGui.QImage

PSNR

An index value that compares PSNR with two images. Currently, SSIM is better than PSNR, but PSNR is also often used. The higher the value, the better the image quality, and the standard quality is PSNR between 30 and 50 when measuring the degree of compression deterioration.

The formula is as follows: MSE is the mean square error and MAX is 255.

PSNR = 10 \times \log 10\frac{MAX^2}{MSE}

The function to find PSNR is as follows. You can find PSNR at high speed by using the ʻImageStat` module.

def psnr(img1, img2):
    diff_img = ImageChops.difference(img1, img2)
    stat = ImageStat.Stat(diff_img)
    mse = sum(stat.sum2) / len(stat.count) / stat.count[0]
    return 10 * math.log10(255 ** 2 / mse)

In addition, it is currently difficult to obtain SSIM at high speed with PIL / Pillow, so it is better to use the pyssim module or OpenCV.

reference

[Peak signal-to-noise ratio-Wikipedia](https://ja.wikipedia.org/wiki/%E3%83%94%E3%83%BC%E3%82%AF%E4%BF%A1%E5%8F % B7% E5% AF% BE% E9% 9B% 91% E9% 9F% B3% E6% AF% 94)

Line art extraction

gray = img.convert("L")
gray2 = gray.filter(ImageFilter.MaxFilter(5))
senga_inv = ImageChops.difference(gray, gray2)
senga = ImageOps.invert(senga_inv)

I referred to the method of here, which is very wonderful.