Author丨Jiefa Shouzhangsheng@Zhihu
Link丨https://zhuanlan.zhihu.com/p/559887437
Using data augmentation techniques can increase the diversity of images in the dataset, thereby improving the model’s performance and generalization ability. The main image augmentation techniques include:
-
Resizing -
Grayscale Transformation -
Normalization -
Random Rotation -
Center Cropping -
Random Cropping -
Gaussian Blur -
Brightness and Contrast Adjustment -
Horizontal Flip -
Vertical Flip -
Gaussian Noise -
Random Blocks -
Central Region
Resizing
Before starting to resize the image, we need to import the data (taking fundus images as an example).
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/000001.tif'))
torch.manual_seed(0) # Set the seed for generating random numbers on CPU for reproducibility
print(np.asarray(orig_img).shape) #(800, 800, 3)
# Resizing the image
resized_imgs = [T.Resize(size=size)(orig_img) for size in [128,256]]
# plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('resize:128*128')
ax2.imshow(resized_imgs[0])
ax3 = plt.subplot(133)
ax3.set_title('resize:256*256')
ax3.imshow(resized_imgs[1])
plt.show()
Grayscale Transformation
This operation converts RGB images to grayscale images.
gray_img = T.Grayscale()(orig_img)
# plt.figure('resize:128*128')
ax1 = plt.subplot(121)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(122)
ax2.set_title('gray')
ax2.imshow(gray_img,cmap='gray')
Normalization
Normalization can speed up the computation speed of models based on neural network structures and accelerate the learning speed.
-
Subtract the channel mean from each input channel -
Divide it by the channel standard deviation.
normalized_img = T.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))(T.ToTensor()(orig_img))
normalized_img = [T.ToPILImage()(normalized_img)]
# plt.figure('resize:128*128')
ax1 = plt.subplot(121)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(122)
ax2.set_title('normalize')
ax2.imshow(normalized_img[0])
plt.show()
Random Rotation
Rotate the image at a designed angle
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
rotated_imgs = [T.RandomRotation(degrees=90)(orig_img)]
print(rotated_imgs)
plt.figure('resize:128*128')
ax1 = plt.subplot(121)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(122)
ax2.set_title('90°')
ax2.imshow(np.array(rotated_imgs[0]))
Center Cropping
Crop the center region of the image
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
center_crops = [T.CenterCrop(size=size)(orig_img) for size in (128,64)]
plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('128*128°')
ax2.imshow(np.array(center_crops[0]))
ax3 = plt.subplot(133)
ax3.set_title('64*64')
ax3.imshow(np.array(center_crops[1]))
plt.show()
Random Cropping
Randomly crop a part of the image
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
random_crops = [T.RandomCrop(size=size)(orig_img) for size in (400,300)]
plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('400*400')
ax2.imshow(np.array(random_crops[0]))
ax3 = plt.subplot(133)
ax3.set_title('300*300')
ax3.imshow(np.array(random_crops[1]))
plt.show()
Gaussian Blur
Use a Gaussian kernel to blur the image
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
blurred_imgs = [T.GaussianBlur(kernel_size=(3, 3), sigma=sigma)(orig_img) for sigma in (3,7)]
plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('sigma=3')
ax2.imshow(np.array(blurred_imgs[0]))
ax3 = plt.subplot(133)
ax3.set_title('sigma=7')
ax3.imshow(np.array(blurred_imgs[1]))
plt.show()
Brightness, Contrast, and Saturation Adjustment
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
# random_crops = [T.RandomCrop(size=size)(orig_img) for size in (832,704, 256)]
colorjitter_img = [T.ColorJitter(brightness=(2,2), contrast=(0.5,0.5), saturation=(0.5,0.5))(orig_img)]
plt.figure('resize:128*128')
ax1 = plt.subplot(121)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(122)
ax2.set_title('colorjitter_img')
ax2.imshow(np.array(colorjitter_img[0]))
plt.show()
Horizontal Flip
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
HorizontalFlip_img = [T.RandomHorizontalFlip(p=1)(orig_img)]
plt.figure('resize:128*128')
ax1 = plt.subplot(121)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(122)
ax2.set_title('colorjitter_img')
ax2.imshow(np.array(HorizontalFlip_img[0]))
plt.show()
Vertical Flip
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
VerticalFlip_img = [T.RandomVerticalFlip(p=1)(orig_img)]
plt.figure('resize:128*128')
ax1 = plt.subplot(121)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(122)
ax2.set_title('VerticalFlip')
ax2.imshow(np.array(VerticalFlip_img[0]))
# ax3 = plt.subplot(133)
# ax3.set_title('sigma=7')
# ax3.imshow(np.array(blurred_imgs[1]))
plt.show()
Gaussian Noise
Add Gaussian noise to the image. The higher the noise factor, the greater the noise in the image.
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
def add_noise(inputs, noise_factor=0.3):
noisy = inputs + torch.randn_like(inputs) * noise_factor
noisy = torch.clip(noisy, 0., 1.)
return noisy
noise_imgs = [add_noise(T.ToTensor()(orig_img), noise_factor) for noise_factor in (0.3, 0.6)]
noise_imgs = [T.ToPILImage()(noise_img) for noise_img in noise_imgs]
plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('noise_factor=0.3')
ax2.imshow(np.array(noise_imgs[0]))
ax3 = plt.subplot(133)
ax3.set_title('noise_factor=0.6')
ax3.imshow(np.array(noise_imgs[1]))
plt.show()
Random Blocks
Randomly apply square patches to the image. The more patches there are, the harder it is for the neural network to solve the problem.
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
def add_random_boxes(img,n_k,size=64):
h,w = size,size
img = np.asarray(img).copy()
img_size = img.shape[1]
boxes = []
for k in range(n_k):
y,x = np.random.randint(0,img_size-w,(2,))
img[y:y+h,x:x+w] = 0
boxes.append((x,y,h,w))
img = Image.fromarray(img.astype('uint8'), 'RGB')
return img
blocks_imgs = [add_random_boxes(orig_img,n_k=10)]
plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('10 black boxes')
ax2.imshow(np.array(blocks_imgs[0]))
plt.show()
Central Region
Similar to random blocks, but patches are added to the center of the image
from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
import sys
import torch
import numpy as np
import torchvision.transforms as T
plt.rcParams["savefig.bbox"] = 'tight'
orig_img = Image.open(Path('image/2.png'))
def add_central_region(img, size=32):
h, w = size, size
img = np.asarray(img).copy()
img_size = img.shape[1]
img[int(img_size / 2 - h):int(img_size / 2 + h), int(img_size / 2 - w):int(img_size / 2 + w)] = 0
img = Image.fromarray(img.astype('uint8'), 'RGB')
return img
central_imgs = [add_central_region(orig_img, size=128)]
plt.figure('resize:128*128')
ax1 = plt.subplot(131)
ax1.set_title('original')
ax1.imshow(orig_img)
ax2 = plt.subplot(132)
ax2.set_title('')
ax2.imshow(np.array(central_imgs[0]))
#
# ax3 = plt.subplot(133)
# ax3.set_title('20 black boxes')
# ax3.imshow(np.array(blocks_imgs[1]))
plt.show()