torchvision.transforms的使用

参考链接: torchvision.transforms

参考链接: 透视变换原理实例代码详解
参考链接: 图像几何变换之透视变换

打开并显示图片文件:

from torchvision import transforms
from matplotlib import pyplot as plt
import numpy as np
import torch
from PIL import Image

img = Image.open("KJ.jpg")
img.show()

plt.imshow(img)
plt.show()

运行结果展示:

PIL image转为tensor类型:

from torchvision import transforms
from matplotlib import pyplot as plt
import numpy as np
import torch
from PIL import Image

img = Image.open("KJ.jpg")

to_tensor = transforms.ToTensor()
img_t = to_tensor(img) # 注意：这里会自动调整通道维度的位置，即通道维度放到最开头
print('img_t.shape:',img_t.shape)  # img_t.shape: torch.Size([3, 374, 500])
print('type(img_t):',type(img_t))  # type(img_t): <class 'torch.Tensor'>
print(img_t.shape, img_t.dtype)  # torch.Size([3, 374, 500]) torch.float32

plt.imshow(img_t.permute(1, 2, 0))  # 需要将通道所对应的维度放到最后
plt.show()

运行结果:

中心裁剪加随机旋转:

from torchvision import transforms
from matplotlib import pyplot as plt
import numpy as np
import torch
from PIL import Image

img = Image.open("KJ.jpg")
trans4cxq = transforms.Compose([
    transforms.CenterCrop((374,400)),
    transforms.RandomRotation([-180,180]),
    transforms.ToTensor(),
    ])

img_tr = trans4cxq(img)
plt.imshow(img_tr.permute(1, 2, 0))  # 需要将通道所对应的维度放到最后
plt.show()

运行结果展示:

透视变换:

from torchvision import transforms
from matplotlib import pyplot as plt
import numpy as np
import torch
from PIL import Image

img = Image.open("KJ.jpg")
scale = 5.0
img = Image.open("poker.png")
img_func = transforms.functional.perspective(
    img, 
    [[63,72], [162,30], [270,144], [151,218]], 
    [[10*scale,20*scale], [30*scale,20*scale], [30*scale,40*scale], [10*scale,40*scale]],
    interpolation=3)

img_func.show()

plt.imshow(img_func)  
plt.show()

所使用的图片和运行结果展示:

使用函数来实现随机旋转,并且添加填充:

from torchvision import transforms
from matplotlib import pyplot as plt
import numpy as np
import torch
from PIL import Image

img = Image.open("KJ.jpg")
img = Image.open("KJ.jpg")
img_func = transforms.functional.rotate(
    img, 45, 
    resample=False, 
    expand=True, 
    center=None)

img_func.show()
plt.imshow(img_func)  
plt.show()

运行效果:

完整实验代码:

from torchvision import transforms
from matplotlib import pyplot as plt
import numpy as np
import torch
from PIL import Image

img = Image.open("KJ.jpg")
# img.show()

# plt.imshow(img)
# plt.show()



# to_tensor = transforms.ToTensor()
# img_t = to_tensor(img) # 注意：这里会自动调整通道维度的位置，即通道维度放到最开头
# print('img_t.shape:',img_t.shape)  # img_t.shape: torch.Size([3, 374, 500])
# print('type(img_t):',type(img_t))  # type(img_t): <class 'torch.Tensor'>
# print(img_t.shape, img_t.dtype)  # torch.Size([3, 374, 500]) torch.float32

# plt.imshow(img_t.permute(1, 2, 0))  # 需要将通道所对应的维度放到最后
# plt.show()



# trans4cxq = transforms.Compose([
#     transforms.CenterCrop((374,400)),
#     transforms.RandomRotation([-180,180]),
#     transforms.ToTensor(),
#     ])

# img_tr = trans4cxq(img)
# plt.imshow(img_tr.permute(1, 2, 0))  # 需要将通道所对应的维度放到最后
# plt.show()





# scale = 5.0
# img = Image.open("poker.png")
# img_func = transforms.functional.perspective(
#     img, 
#     [[63,72], [162,30], [270,144], [151,218]], 
#     [[10*scale,20*scale], [30*scale,20*scale], [30*scale,40*scale], [10*scale,40*scale]],
#     interpolation=3)

# img_func.show()

# plt.imshow(img_func)  
# plt.show()






# scale = 100.0
# img = Image.open("cxq.jpg")
# img_func = transforms.functional.perspective(
#     img, 
#     [[3097,3015], [1649,3405], [591,1393], [1761,1477]], 
#     [[10*scale,20*scale], [30*scale,20*scale], [30*scale,40*scale], [10*scale,40*scale]],
#     interpolation=3)

# img_func.show()





img = Image.open("KJ.jpg")
img_func = transforms.functional.rotate(
    img, 45, 
    resample=False, 
    expand=True, 
    center=None)

img_func.show()
plt.imshow(img_func)  
plt.show()

实验所使用的图片文件: