如何为数据增强创建噪声图像

Posted 2023-02-16

【中文标题】如何为数据增强创建噪声图像

【英文标题】：How to create noisy images for data augmentation

【发布时间】：2020-05-01 07:15:24

【问题描述】：

我关注了有关向图像添加噪点的问题的最受好评的答案。但是它对我不起作用。我只想在使用 Python 时观察图像上的不同噪声效果 How to add noise (Gaussian/salt and pepper etc) to image in Python with OpenCV

据我所知，图像是 uint8 类型的吗？我不确定这种类型是否可以带小数。

盐和胡椒部分也不起作用

from numpy import shape, asarray
import numpy as np
import cv2
from PIL import Image
def noisy(noise_typ,image):
    if noise_typ == "gauss":
      row,col,ch= image.shape
      mean = 0
      var = 0.1
      sigma = var**0.5
      gauss = np.random.normal(mean,sigma,(row,col,ch))
      gauss = gauss.reshape(row,col,ch)
      noisy = image + gauss
      return noisy
    elif noise_typ == "s&p":
      row,col,ch = image.shape
      s_vs_p = 0.5
      amount = 0.004
      out = np.copy(image)
      # Salt mode
      num_salt = np.ceil(amount * image.size * s_vs_p)
      coords = [np.random.randint(0, i - 1, int(num_salt))
              for i in image.shape]
      out[coords] = 1

      # Pepper mode
      num_pepper = np.ceil(amount* image.size * (1. - s_vs_p))
      coords = [np.random.randint(0, i - 1, int(num_pepper))
              for i in image.shape]
      out[coords] = 0
      return out
    elif noise_typ == "poisson":
      vals = len(np.unique(image))
      vals = 2 ** np.ceil(np.log2(vals))
      noisy = np.random.poisson(image * vals) / float(vals)
      return noisy
    elif noise_typ =="speckle":
      row,col,ch = image.shape
      gauss = np.random.randn(row,col,ch)
      gauss = gauss.reshape(row,col,ch)        
      noisy = image + image * gauss
      return noisy

pic = Image.open('obamaface1.jpg')
pic = pic.convert('RGB')
pixels = asarray(pic)
image = Image.fromarray(pixels)

target = noisy('speckle', pixels)
target = Image.fromarray(target)
print(target)

最后一行是查看终端输出的内容。它的输出是

  File "C:\Users\Jerome Ariola\AppData\Local\Programs\Python\Python36\lib\site-packages\PIL\Image.py", line 2647, in fromarray
    raise TypeError("Cannot handle this data type")
TypeError: Cannot handle this data type

评论target = Image.fromarray(target)给我：

[[[ 10.03013352   7.53745105  11.03977542]
  [  9.38952149   7.81507808  11.53212491]
  [  9.76439692   7.88213107  11.47620008]
  ...
  [  6.76471119   5.09559321   5.9144036 ]
  [  7.34123162   4.92342273   6.31726796]
  [255.13791218 253.89755922 255.15403824]]

 [[  9.90775807   8.49642977  10.86023707]
  [  9.71078442   7.94264649  11.18820572]
  [  9.91127254   8.15716707  11.04770154]
  ...
  [  7.05173864   4.89094663   5.67662439]
  [  7.10166986   5.47480635   6.11892638]
  [255.03879603 254.07485578 254.88072098]]

 [[  9.81995678   7.55439474  11.08609859]
  [ 10.32135236   7.5301714   11.03612056]
  [ 10.17215819   8.09537629  11.30984933]
  ...
  [  7.13999574   5.12009845   7.8678079 ]
  [  7.31635614   5.1527127    8.23318054]
  [255.12283461 254.01880276 254.76894074]]

 ...

 [[ 19.72596723  22.29694693  20.95524912]
  [ 19.30898519  21.61944993  20.85653566]
  [ 20.45174165  20.55101246  21.1739277 ]
  ...
  [ 13.89796331  11.73865315  12.50874487]
  [ 14.13985843  11.97177032  12.80855176]
  [255.04963076 254.23626115 254.75904336]]

 [[ 19.17915912  21.2224852   18.37260714]
  [ 19.1068802   20.2797369   17.96846182]
  [ 20.37263348  20.23856465  18.02893703]
  ...
  [ 14.48307596  12.46348446  15.43437954]
  [ 14.11840104  12.35783324  14.64863437]
  [254.99657596 253.95241488 255.34200558]]

 [[ 20.03354477  22.02402748  18.45595882]
  [ 19.74202893  22.59472663  19.19910502]
  [ 21.96931817  22.22425014  19.59694792]
  ...
  [ 14.68256917  12.37915145  15.07832362]
  [ 14.25010143  12.45545202  14.65549651]
  [254.96105357 254.17655349 255.29863654]]]

我想我也会重写它，例如从image.shape 到np.shape(image)

def noisify(type, target):
    if type == 'gauss':
        row,col, ch = np.shape(target)
        mean = 0
        var = 0.1
        sigma = var**0.5
        gauss = np.random.normal(mean, sigma, (row,col,ch))
        noise = target + gauss
        return noise
    elif type == 'sap':
        row, col, ch = np.shape(target)
        s_vs_p = 0.5
        amount = 0.004
        out = np.copy(target)
        #salt
        num_salt = np.ceil(amount * np.size(target) * s_vs_p)
        coords = [np.random.randint(0,i-1, int(num_salt))
                for i in np.shape(target)]
        out[coords] = 1
        #pepper
        num_pepper = np.ceil(amount* np.size(target) * (1. - s_vs_p))
        coords = [np.random.randint(0,i-1, int(num_pepper))
                for i in np.shape(target)]
        out[coords] = 0
        return out
    elif type == 'poisson':
        vals = len(np.unique(target))
        vals = 2 ** np.ceil(np.log2(vals))
        noise = np.random.poisson(target * vals) / float(vals)
        return noise
    elif type == 'speckle':
        row, col, ch = np.shape(target)
        gauss = np.random.randn(row,col,ch)
        gauss = np.reshape(gauss,(row,col,ch))
        noise = target + target * gauss
        return noise

任何帮助将不胜感激。

【问题讨论】：

你能定义什么不工作吗？它在视觉上看起来是否相同，或者每个像素的 RGB 是否保持相同？

在添加噪点之前，您是否将输入图像转换为浮动图像。

@scottsaenz for 'gauss' 似乎可以完成这项工作。但是我无法创建一个允许我查看它的外观的 PIL 图像（通过 Image.fromarray() 应该在临时文件夹上为我创建一个图像。不起作用的是，也许 PIL 图像不能用十进制值制作，但是还能怎么添加噪音呢？

@fmw42 没有？我没有。我拍摄了图像并使用 asarray 将其从 PIL 图像转换为 uint8 张量或其他东西。我会考虑使用花车...

查看您从中获取代码的参考。在顶部它说将输入转换为浮点数。

【参考方案1】：

这是使用 OpenCV + skimage.util.random_noise 的矢量化方法。您可以尝试使用localvar、pepper、s&p 和speckle 等噪声模式以获得所需的结果。您可以使用amount 参数设置噪声的比例。这是一个使用s&p 和amount=0.011 的示例：

输入图像

结果

与amount=0.051:

import cv2
import numpy as np
from skimage.util import random_noise

# Load the image
image = cv2.imread('1.png', 0)

# Add salt-and-pepper noise to the image
noise = random_noise(image, mode='s&p', amount=0.011)

# The above function returns a floating-point image in the range [0, 1]
# so need to change it to 'uint8' with range [0,255]
noise = np.array(255 * noise, dtype=np.uint8)

cv2.imshow('noise',noise)
cv2.imwrite('noise.png',noise)
cv2.waitKey()

【讨论】：

这太棒了！非常感谢，它完全实现了我想要的。虽然有一个图书馆，但我从失败的经历中学到了。谢谢