在Pytorch创建数据集时,常常会有transform.ToPILImage
和transform.ToTensor
两个函数,前一个函数是将numpy转变为PILImage形式,第二个函数是将PILImage形式转变为tensor形式方便计算,转换时需要注意以下几点
图片一共有三种形式,PILImage形式,tensor形式以及numpy形式
- numpy必须为uint8形式,范围为[0, 255]
- tensor必须为FloatTensor形式,范围为[0, 1]
- tensor如果要转换为numpy需要乘上255,最好归一化
- tensor的数据形式为CHW,numpy则为HWC,使用
array.transpose(1, 2, 0)
进行转化 - PILImage以及tensor的通道形式均为RGB,而opencv(numpy)的则为BGR,转化时需要用
cv2.cvtColor(img,cv2.COLOR_xxx)
来进行转化,否则图片颜色会变化 - PILImage形式的图片显示时用
img.show()
即可;numpy形式可用opencv进行显示,cv2.imshow("image", array)
,而tensor形式的图片则需要转化为numpy用opencv显示(可见下面代码)
import cv2
import torchvision.transforms as transforms
import numpy as np
import torch
import matplotlib.pyplot as plt
from PIL import Image
'''PIL image'''
image = Image.open("./food-11/0000.jpg")
print(image) # <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=384x512 at 0x7F1CFE47DD60>
image2 = transforms.ToTensor()(image) # to simplify computation
print(image2) # tensor.data[0, 1]
print(image2.dtype) # tensor.float32
print(image2.shape) # tensor.Size([3, 512, 384])
# if you want opencv to support tensor image, you should know below
# opencv support numpy format, dtype is uint8, pixle range [0, 255]
# when use ToTensor, tensor range is [0, 1], dtype is FloatTensor
# channel: PIL && torch:RGB opencv:BGR
# dimensional: torch: CHW numpy: HWC
# show tensor
tensor_ = image2
array_ = tensor_.numpy() # convert tensor to numpy
max = array_.max()
array_ = array_*255/max # expand pixel from [0, 1] to [0, 255], then normalization
array_convert = np.uint8(array_) # opencv only support uint8 numpy.array
print("array_convert:", array_convert.shape) # array_convert: (3, 512, 384)
array_convert = array_convert.transpose(1, 2, 0) # array_convert: (512, 384, 3)
cv2.imshow("image", array_convert) # different from the original image
cv2.waitKey()
array_convert = cv2.cvtColor(array_convert, cv2.COLOR_RGB2BGR) # convert RGB to BGR
cv2.imshow("image2", array_convert) # the same to the original image
cv2.waitKey()
'''opencv'''
img = cv2.imread("./food-11/0000.jpg")
print(img) # np.array.data[0, 255]
print(type(img)) # <class 'numpy.ndarray'>
print(img.dtype) # uint8
print(img.shape) # (512, 384, 3)
cv2.imshow("img", img)
cv2.waitKey()
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # if don't convert, the image will distort
img2 = transforms.ToPILImage()(img)
img2.show()
如果图片为numpy类型的浮点数float32,可以直接用skimage.img_as_ubyte将其转换为255范围的uint8类型,img=skimage.img_as_ubyte(img)