PyTorch学习笔记——图像处理（transforms.Normalize 归一化）

回顾 torchvision.ToTensor

在看这一片博客之前，需要先浏览以下我的上一篇博客
PyTorch学习笔记——图像处理（torchvision.ToTensor）

import torchvision.transforms as transforms
import numpy as np
import torch
import cv2
from matplotlib import pyplot as plt
import matplotlib.image as imgplt

img_path = "1.jpg"

# transforms.ToTensor()
transform1 = transforms.Compose([
    transforms.ToTensor(), # range [0, 255] -> [0.0,1.0]
    ]
)

##numpy.ndarray
img = cv2.imread(img_path)# 读取图像
img1 = transform1(img) # 归一化到 [0.0,1.0]
print("img.shape = ",img.shape)
print("img1.shape = ",img1.shape)

print("img = ",img)
print("img1 = ",img1)

img.shape =  (424, 640, 3)
img1.shape =  torch.Size([3, 424, 640])
img =  [[[249 246 231]
  [248 246 228]
  [247 245 227]
  ...
  [ 17  19  20]
  [ 20  19  21]
  [ 32  14  21]]

 [[249 246 231]
  [248 246 228]
  [247 245 227]
  ...
  [ 29  30  20]
  [ 29  30  20]
  [ 15  24  14]]

 [[248 245 230]
  [248 245 230]
  [248 246 228]
  ...
  [ 80  41  26]
  [ 80  41  26]
  [ 45  36  16]]

 ...

 [[169 175 216]
  [174 178 219]
  [167 173 210]
  ...
  [ 14  17  22]
  [ 14  17  22]
  [ 15  18  23]]

 [[151 157 202]
  [163 167 208]
  [172 174 209]
  ...
  [ 14  17  22]
  [ 13  16  21]
  [ 13  16  21]]

 [[168 171 215]
  [173 174 212]
  [179 177 207]
  ...
  [ 13  16  21]
  [ 12  15  20]
  [ 11  13  21]]]
img1 =  tensor([[[0.9765, 0.9725, 0.9686,  ..., 0.0667, 0.0784, 0.1255],
         [0.9765, 0.9725, 0.9686,  ..., 0.1137, 0.1137, 0.0588],
         [0.9725, 0.9725, 0.9725,  ..., 0.3137, 0.3137, 0.1765],
         ...,
         [0.6627, 0.6824, 0.6549,  ..., 0.0549, 0.0549, 0.0588],
         [0.5922, 0.6392, 0.6745,  ..., 0.0549, 0.0510, 0.0510],
         [0.6588, 0.6784, 0.7020,  ..., 0.0510, 0.0471, 0.0431]],

        [[0.9647, 0.9647, 0.9608,  ..., 0.0745, 0.0745, 0.0549],
         [0.9647, 0.9647, 0.9608,  ..., 0.1176, 0.1176, 0.0941],
         [0.9608, 0.9608, 0.9647,  ..., 0.1608, 0.1608, 0.1412],
         ...,
         [0.6863, 0.6980, 0.6784,  ..., 0.0667, 0.0667, 0.0706],
         [0.6157, 0.6549, 0.6824,  ..., 0.0667, 0.0627, 0.0627],
         [0.6706, 0.6824, 0.6941,  ..., 0.0627, 0.0588, 0.0510]],

        [[0.9059, 0.8941, 0.8902,  ..., 0.0784, 0.0824, 0.0824],
         [0.9059, 0.8941, 0.8902,  ..., 0.0784, 0.0784, 0.0549],
         [0.9020, 0.9020, 0.8941,  ..., 0.1020, 0.1020, 0.0627],
         ...,
         [0.8471, 0.8588, 0.8235,  ..., 0.0863, 0.0863, 0.0902],
         [0.7922, 0.8157, 0.8196,  ..., 0.0863, 0.0824, 0.0824],
         [0.8431, 0.8314, 0.8118,  ..., 0.0824, 0.0784, 0.0824]]])

这是上一篇的代码，可以看到，使用transforms.ToTensor将图像转化为tensor

归一化 transforms.Normalize

transform1 = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean = (0.5, 0.5, 0.5), std = (0.5, 0.5, 0.5))
    ]
)

将源代码修改为现在这样

img_path = "1.jpg"

# transforms.ToTensor()
transform2 = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean = (0.5, 0.5, 0.5), std = (0.5, 0.5, 0.5))
    ]
)

##numpy.ndarray
import torchvision.transforms as transforms
import numpy as np
import torch
import cv2
from matplotlib import pyplot as plt
import matplotlib.image as imgplt
img_path = "1.jpg"

# transforms.ToTensor()
transform1 = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean = (0.5, 0.5, 0.5), std = (0.5, 0.5, 0.5))
    ]
)

##numpy.ndarray
img = cv2.imread(img_path)# 读取图像
img1 = transform1(img) # 归一化到 [0.0,1.0]
print("img.shape = ",img.shape)
print("img1.shape = ",img1.shape)

print("img = ",img)
print("img1 = ",img1)

img.shape =  (424, 640, 3)
img1.shape =  torch.Size([3, 424, 640])
img =  [[[249 246 231]
  [248 246 228]
  [247 245 227]
  ...
  [ 17  19  20]
  [ 20  19  21]
  [ 32  14  21]]

 [[249 246 231]
  [248 246 228]
  [247 245 227]
  ...
  [ 29  30  20]
  [ 29  30  20]
  [ 15  24  14]]

 [[248 245 230]
  [248 245 230]
  [248 246 228]
  ...
  [ 80  41  26]
  [ 80  41  26]
  [ 45  36  16]]

 ...

 [[169 175 216]
  [174 178 219]
  [167 173 210]
  ...
  [ 14  17  22]
  [ 14  17  22]
  [ 15  18  23]]

 [[151 157 202]
  [163 167 208]
  [172 174 209]
  ...
  [ 14  17  22]
  [ 13  16  21]
  [ 13  16  21]]

 [[168 171 215]
  [173 174 212]
  [179 177 207]
  ...
  [ 13  16  21]
  [ 12  15  20]
  [ 11  13  21]]]
img1 =  tensor([[[ 0.9529,  0.9451,  0.9373,  ..., -0.8667, -0.8431, -0.7490],
         [ 0.9529,  0.9451,  0.9373,  ..., -0.7725, -0.7725, -0.8824],
         [ 0.9451,  0.9451,  0.9451,  ..., -0.3725, -0.3725, -0.6471],
         ...,
         [ 0.3255,  0.3647,  0.3098,  ..., -0.8902, -0.8902, -0.8824],
         [ 0.1843,  0.2784,  0.3490,  ..., -0.8902, -0.8980, -0.8980],
         [ 0.3176,  0.3569,  0.4039,  ..., -0.8980, -0.9059, -0.9137]],

        [[ 0.9294,  0.9294,  0.9216,  ..., -0.8510, -0.8510, -0.8902],
         [ 0.9294,  0.9294,  0.9216,  ..., -0.7647, -0.7647, -0.8118],
         [ 0.9216,  0.9216,  0.9294,  ..., -0.6784, -0.6784, -0.7176],
         ...,
         [ 0.3725,  0.3961,  0.3569,  ..., -0.8667, -0.8667, -0.8588],
         [ 0.2314,  0.3098,  0.3647,  ..., -0.8667, -0.8745, -0.8745],
         [ 0.3412,  0.3647,  0.3882,  ..., -0.8745, -0.8824, -0.8980]],

        [[ 0.8118,  0.7882,  0.7804,  ..., -0.8431, -0.8353, -0.8353],
         [ 0.8118,  0.7882,  0.7804,  ..., -0.8431, -0.8431, -0.8902],
         [ 0.8039,  0.8039,  0.7882,  ..., -0.7961, -0.7961, -0.8745],
         ...,
         [ 0.6941,  0.7176,  0.6471,  ..., -0.8275, -0.8275, -0.8196],
         [ 0.5843,  0.6314,  0.6392,  ..., -0.8275, -0.8353, -0.8353],
         [ 0.6863,  0.6627,  0.6235,  ..., -0.8353, -0.8431, -0.8353]]])

可以看到，最后的数组，值的范围变成了[-1,1]

公式

transforms.Normalize使用如下公式进行归一化：

channel=（channel-mean）/std
也就是说 ( (0,1) - 0.5 ) / 0.5 = (-1,1)