Read digital image processing
I am studying with a book called Digital Image Processing. I implemented (contrast conversion) for the purpose of deepening understanding.
* Brightness contrast conversion *
Contrast is the degree of change in brightness and color. (Personal interpretation) Strong contrast → Easy to understand brightness and color transfer, easy to understand the outline of an object. Contrast is weak → It is difficult to understand the brightness and color transfer, and it is difficult to understand the outline of an object.
* Tone curve *
By changing the relationship between input and output, the color tone of the image can be changed. Therefore, it becomes easy to see or difficult to see.
* The first piece to use this time *
Use a dark sample like the one in the picture below to make it easier to see the effect.
Features: Pretty dark
python
#Required library
import cv2
import numpy as np
import matplotlib.pyplot as plt
#Photo path
pic_path = 'pet-bottle-pic/test/camera_pic/aquari3.jpg'
#Read
pic_image = cv2.imread(pic_path)
#Convert to gray
pic_image = cv2.cvtColor(pic_image, cv2.COLOR_BGR2GRAY)
#When read with cv, plt.cmap when using imshow='gray'Note that it will not be gray unless you specify
plt.imshow(pic_image, cmap='gray')
plt.title('sample_dark')
plt.savefig("aquari_dark.jpg ")
↓ It's dark and almost invisible
python
#Pixel
pic_image.shape # (2464, 3280)
Pixel value distribution
python
#Flatten in one dimension to make a histogram
flat_data = np.ravel(pic_image)
#Bar span
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_dark')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_dark.png ")
↓ There is a bias on the black side.
Implementation of contrast conversion
* Line type *
Depending on how it is made, the contrast in any range can be increased. (Although it is a little unnatural) Since the sample image is dense in the range of pixel values 0 to 50, the angle was set in that range.
python
def bend_line(X):
y = X*5
y = np.where(y>=255, 255, y)
return y
Apply the function that outputs as shown in the graph below.
python
X = np.arange(0, 255, 1)
y = bend_line(X)
plt.title('sample_dark_bend_line')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
plt.plot(X, y)
plt.savefig("bend_line.png ")
Photo after application
python
pic_func1 = bend_line(pic_image)
plt.imshow(pic_func1, cmap='gray')
plt.title('sample_dark_after_bend_line')
plt.savefig("aquari_dark_after_bend_line.jpg ")
↓ You can now see PET bottles. (Although it looks a little dirty)
Let's look at the distribution.
python
flat_data = np.ravel(pic_func1)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_dark_after_bend_line')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_dark_after_bend_line.png ")
↓ The distribution has been smoothed out a little.
* Exponentiation tone curve (gamma correction) *
Suitable for correcting images that are too dark or too bright to understand. It is suitable for increasing the contrast of images with a distribution near the edges.
Implement the following formula called the gamma function.
python
def g_function(X, gamma):
y = 255*(X/255)**(1/gamma)
return y
Graph for each γ value
python
X = np.arange(0, 255, 1)
gamma = [3, 2, 1.5, 1, 0.5, 0.33]
labels = ["3", "2", "1.5", "1", "0.5", "0.33"]
plt.title('g_function')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
for g, l in zip(gamma, labels):
y = g_function(X, g)
plt.plot(X, y, label = l)
plt.legend()
plt.savefig("g_function.png ")
Large γ → Increases the contrast of dark images overall. γ is small → Increases the contrast of bright images overall.
python
gamma = 3
pic_gfunc = g_function(pic_image, gamma)
plt.imshow(pic_gfunc, cmap='gray')
plt.title('sample_dark_after_gfunc')
plt.savefig("aquari_dark_after_gfunc.jpg ")
↓ It's pretty clean.
python
flat_data = np.ravel(pic_gfunc)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_dark_after_gfunc')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_dark_after_gfunc.png ")
↓ Similarly, the distribution is smoothed.
* Next image to use
Features: Normal
python
#Photo path
pic_path = 'pet-bottle-pic/test/camera_pic/aquari.jpg'
#Read
pic_image = cv2.imread(pic_path)
#Convert to gray
pic_image = cv2.cvtColor(pic_image, cv2.COLOR_BGR2GRAY)
#When read with cv, plt.cmap when using imshow='gray'Note that it will not be gray unless you specify
plt.imshow(pic_image, cmap='gray')
plt.title('sample_normal')
plt.savefig("aquari.jpg ")
Pixel value distribution
python
#Flatten in one dimension to make a histogram
flat_data = np.ravel(pic_image)
#Bar span
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_normal')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari.png ")
* S-shaped curve type *
It is suitable for increasing the contrast of images with a distribution near the center.
Implement the following formula
python
def s_func(X, a):
y = 255/(1 + np.exp(a*(-1*X + 127)))
return y
python
X = np.arange(0, 255, 1)
a = 0.05
y = s_func(X, a)
plt.title('s_function')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
plt.plot(X, y)
plt.savefig("s_function.png ")
python
a = 0.05
y = s_func(pic_image, a)
plt.imshow(y, cmap='gray')
plt.title('sample_after_sfunc')
plt.savefig("aquari_after_sfunc.jpg ")
↓ It was a little clear.
python
flat_data = np.ravel(y)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_after_sfunc')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_sfunc.png ")
↓ The histogram became flat.
* Special effects *
There are some mysterious aspects to its usage, but I implemented it.
* Invert *
Bright → dark Dark → bright become.
python
def rev_func(X):
y = -X + 255
return y
python
X = np.arange(0, 255, 1)
y = rev_func(X)
plt.title('Rev_function')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
plt.plot(X, y)
plt.savefig("rev_func.png ")
python
y = rev_func(pic_image)
plt.imshow(y, cmap='gray')
plt.title('sample_after_rev_func')
plt.savefig("aquari_after_rev_func.jpg ")
python
flat_data = np.ravel(y)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_after_rev_func')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_rev_func.png ")
↓ The distribution is also reversed on the left and right.
* Posterization *
It looks like a poster.
Brightness divided into 4 stages
Interval: 64 First stage: 64 Second stage: 128 Third stage: 192 Fourth stage: 255
python
def post_func(X):
y = np.where(X>=192, 255, X)
y = np.where((y >=128 ) & (y < 192), 170, y)
y = np.where((y >=64 ) & (y < 128), 85, y)
y = np.where(y<64, 0, y)
return y
python
X = np.arange(0, 255, 1)
y = post_func(X)
plt.title('post_function')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
plt.plot(X, y)
plt.savefig("post_func.png ")
python
y = post_func(pic_image)
plt.imshow(y, cmap='gray')
plt.title('sample_after_after_postfunc')
plt.savefig("aquari_after_postfunc.jpg ")
↓ Poster-like
python
flat_data = np.ravel(y)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_after_postfunc')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_postfunc.png ")
* Binarization *
Black and white clear]
python
def to_func(X):
y = np.where(X>=127, 255, X)
y = np.where(y<127, 0, y)
return y
python
X = np.arange(0, 255, 1)
y = to_func(X)
plt.title('2_function')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
plt.plot(X, y)
plt.savefig("2_func.png ")
python
y = to_func(pic_image)
plt.imshow(y, cmap='gray')
plt.title('sample_after_after_tofunc')
plt.savefig("aquari_after_tofunc.jpg ")
↓ Is it hard to see on the contrary?
python
flat_data = np.ravel(y)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_after_totfunc')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_tofunc.png ")
↓ It is binarized.
* Solarization *
It feels like a film when developing a photo.
python
from math import pi
def sora_func(X):
X = (3*pi/255)*X
y = -127.5*(np.cos(X))+127.5
return y
python
X = np.arange(0, 255, 1)
y = sora_func(X)
plt.title('sora_function')
plt.xlabel('INPUT')
plt.ylabel('OUTPUT')
plt.plot(X, y)
plt.savefig("sora_func.png ")
python
y = sora_func(pic_image)
plt.imshow(y, cmap='gray')
plt.title('sample_after_sorafunc')
plt.savefig("aquari_after_sorafunc.jpg ")
↓ Somehow it looks like that
python
flat_data = np.ravel(y)
bins_range = range(0, 260, 5)
#Output histogram
plt.hist(flat_data, bins = bins_range)
plt.title('sample_after_soratfunc')
plt.xlabel('pixel_value')
plt.ylabel('frequency')
plt.savefig("hist_aquari_sorafunc.png ")
End
Recommended Posts