实验需要,将RGB转化为gray,可以单独处理RGB数据,或者在训练时,数据准备阶段,利用transforms直接转化。
RGB转gray可以通过PIL或者CV2相应算法库进行转化。另一种选择是通过公式Y = 0.299 R + 0.587 G + 0.114 B.
从RGB计算灰度(Y)通道 切片数组并转换为一个通道。
class Gray(object)
def __call__(self, tensor):
# TODO: make efficient
R = tensor[0]
G = tensor[1]
B = tensor[2]
tensor[0]=0.299*R+0.587*G+0.114*B
tensor = tensor[0]
tensor = tensor.view(1,w,h)
return tensor
注意PIL读取图片为RGB格式,cv2.imread读取图片为BGR格式。
数据准备阶段 transform转化
import torch
import torchvision
import torchvision.transforms as transforms
from PIL import Image
image_path = "/home/WuHF/whf/pytorch_code/dttNet/dataset/fusion_datasets/lytro-01-A.jpg"
image = Image.open(image_path)
input_transform = transforms.Compose([
transforms.Grayscale(num_output_channels=1), #这一句就是转为单通道灰度图像
transforms.ToTensor(),
])
image_tensor = input_transform(image)
数据单独转化
采用convert函数将RGB转化为Gray,采用expand_dims和repeat函数扩充到三通道。
import numpy as np
from PIL import Image
img_path = 'test.jpg'
img = Image.open(img_path)
# print(np.array(img).shape)
# print(np.array(img)[50:52,50:60,:])
# 转化为灰度图,这里建议先转化为灰度图,再进行resize。
img = img.convert('L')
img = img.resize((144, 288), Image.ANTIALIAS)
# print(np.array(img).shape)
# print(np.array(img)[50:52,50:60])
#灰度图单通道复制扩充到三通道
img = np.expand_dims(img, axis=0)
img = np.array(img).repeat(3, axis=0)
# print(np.array(img).shape)
# print(np.array(img)[...,50:52,50:60])
repeat函数使用
numpy.repeat(a,repeats,axis=None)或者是object(ndarray).repeat(repeats,axis=None)
axis=None,时候就会flatten当前矩阵,实际上就是变成了一个行向量
axis对应维度扩充repeats倍,
repeats可为数,也可为矩阵(针对axis维度的元素),复制repeats次。
a=numpy.array(([[1,2], [3,4]]))
#[8]
numpy.repeat(a,2)
# array([[1,1,2,2,3,3,4,4]])
#0维度2倍 [4,2]
numpy.repeat(a,2,axis=0)
# array([[1,2],
[1,2],
[3,4],
[3,4]])
#1维度2倍 [2,4]
numpy.repeat(a,2,axis=1)
# array([[1,1,2,2],
[3,3,4,4]])
# [5,2] 矩阵对应元素复制n倍
numpy.repeat(a,[2,3],axis=0)
# array([[1,2],
[1,2],
[3,4],
[3,4],
[3,4]])
Numpy扩充矩阵维度(np.expand_dims, np.newaxis)和删除维度(np.squeeze)
np.newaxis和np.expand_dims扩充矩阵维度
import numpy as np
x = np.arange(8).reshape(2, 4)
# 添加第0维,输出shape -> (1, 2, 4),这里np.newaxis所在的位置就是新添加的维度位置
x1 = x[np.newaxis, :]
print(x1.shape)
# 添加第1维, 输出shape -> (2, 1, 4),这里直接指定axis即可指定维度
x2 = np.expand_dims(x, axis=1)
print(x2.shape)
np.squeeze删除矩阵中维度大小为1的维度
"""
squeeze 函数:从数组的形状中删除单维度条目,即把shape中为1的维度去掉
用法:numpy.squeeze(a,axis = None)
1)a表示输入的数组;
2)axis用于指定需要删除的维度,但是指定的维度必须为单维度,否则将会报错;
3)axis的取值可为None 或 int 或 tuple of ints, 可选。若axis为空,则删除所有单维度的条目;
4)返回值:数组
5) 不会修改原数组;
"""
import numpy as np
x = np.arange(10).reshape(1, 1, 10, 1)
print(x.shape,x)
(1, 1, 10, 1)
[[[[0]
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]]]]
#去掉axis=0这个维度
x_squeeze_0 = np.squeeze(x, axis=0)
print(x_squeeze_0.shape, x_squeeze_0)
(1, 10, 1) [[[0]
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]]]
#去掉axis=3这个维度
x_squeeze_3 = np.squeeze(x, axis=3)
print(x_squeeze_3.shape, x_squeeze_3)
(1, 1, 10) [[[0 1 2 3 4 5 6 7 8 9]]]
#去掉axis=0, axis=1这两个维度
x_squeeze_0_1 = np.squeeze(x, axis=(0, 1))
print(x_squeeze_0_1.shape, x_squeeze_0_1)
(10, 1) [[0]
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]]
#去掉所有1维的维度
x_squeeze = np.squeeze(x)
print(x_squeeze.shape, x_squeeze)
(10,) [0 1 2 3 4 5 6 7 8 9]
#去掉不是1维的维度,抛异常
try:
x_squeeze = np.squeeze(x, axis=2)
print(x_squeeze.shape, x_squeeze)
except Exception as e:
print(e)
#cannot select an axis to squeeze out which has size not equal to one