关键词

深度学习中的Data Augmentation方法(转)基于keras

在深度学习中,当数据量不够大时候,常常采用下面4中方法:

1. 人工增加训练集的大小. 通过平移, 翻转, 加噪声等方法从已有数据中创造出一批"新"的数据.也就是Data Augmentation

2. Regularization. 数据量比较小会导致模型过拟合, 使得训练误差很小而测试误差特别大. 通过在Loss Function 后面加上正则项可以抑制过拟合的产生. 缺点是引入了一个需要手动调整的hyper-parameter. 详见 https://www.wikiwand.com/en/Regularization_(mathematics)

3. Dropout. 这也是一种正则化手段. 不过跟以上不同的是它通过随机将部分神经元的输出置零来实现. 详见 http://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf

4. Unsupervised Pre-training. 用Auto-Encoder或者RBM的卷积形式一层一层地做无监督预训练, 最后加上分类层做有监督的Fine-Tuning. 参考 http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.207.1102&rep=rep1&type=pdf

 

下面我们来讨论Data Augmentation:

不同的任务背景下, 我们可以通过图像的几何变换, 使用以下一种或多种组合数据增强变换来增加输入数据的量. 这里具体的方法都来自数字图像处理的内容, 相关的知识点介绍, 网上都有, 就不一一介绍了.

  • 旋转 | 反射变换(Rotation/reflection): 随机旋转图像一定角度; 改变图像内容的朝向;
  • 翻转变换(flip): 沿着水平或者垂直方向翻转图像;
  • 缩放变换(zoom): 按照一定的比例放大或者缩小图像;
  • 平移变换(shift): 在图像平面上对图像以一定方式进行平移; 
    可以采用随机或人为定义的方式指定平移范围和平移步长, 沿水平或竖直方向进行平移. 改变图像内容的位置;
  • 尺度变换(scale): 对图像按照指定的尺度因子, 进行放大或缩小; 或者参照SIFT特征提取思想, 利用指定的尺度因子对图像滤波构造尺度空间. 改变图像内容的大小或模糊程度;
  • 对比度变换(contrast): 在图像的HSV颜色空间,改变饱和度S和V亮度分量,保持色调H不变. 对每个像素的S和V分量进行指数运算(指数因子在0.25到4之间), 增加光照变化;
  • 噪声扰动(noise): 对图像的每个像素RGB进行随机扰动, 常用的噪声模式是椒盐噪声和高斯噪声;
  • 颜色变换(color): 在训练集像素值的RGB颜色空间进行PCA, 得到RGB空间的3个主方向向量,3个特征值, p1, p2, p3, λ1, λ2, λ3. 对每幅图像的每个像素T进行加上如下的变化:

                                        T

      .

代码实现

作为实现部分, 这里介绍一下在python 环境下, 利用已有的开源代码库Keras作为实践:

 1 # -*- coding: utf-8 -*-
 2 __author__ = 'Administrator'
 3 
 4 # import packages
 5 from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
 6 
 7 datagen = ImageDataGenerator(
 8         rotation_range=0.2,
 9         width_shift_range=0.2,
10         height_shift_range=0.2,
11         shear_range=0.2,
12         zoom_range=0.2,
13         horizontal_flip=True,
14         fill_mode='nearest')
15 
16 img = load_img('C:UsersAdministratorDesktopdataAlena.jpg')  # this is a PIL image, please replace to your own file path
17 x = img_to_array(img)  # this is a Numpy array with shape (3, 150, 150)
18 x = x.reshape((1,) + x.shape)  # this is a Numpy array with shape (1, 3, 150, 150)
19 
20 # the .flow() command below generates batches of randomly transformed images
21 # and saves the results to the `preview/` directory
22 
23 i = 0
24 for batch in datagen.flow(x,
25                           batch_size=1,
26                           save_to_dir='C:UsersAdministratorDesktopdataApre',#生成后的图像保存路径
27                           save_prefix='lena',
28                           save_format='jpg'):
29     i += 1
30     if i > 20: #这个20指出要扩增多少个数据
31         break  # otherwise the generator would loop indefinitely

 

主要函数:ImageDataGenerator 实现了大多数上文中提到的图像几何变换方法.

  • rotation_range: 旋转范围, 随机旋转(0-180)度;
  • width_shift and height_shift: 随机沿着水平或者垂直方向,以图像的长宽小部分百分比为变化范围进行平移;
  • rescale: 对图像按照指定的尺度因子, 进行放大或缩小, 设置值在0 - 1之间,通常为1 / 255;
  • shear_range: 水平或垂直投影变换, 参考这里 https://keras.io/preprocessing/image/
  • zoom_range: 按比例随机缩放图像尺寸;
  • horizontal_flip: 水平翻转图像;
  • fill_mode: 填充像素, 出现在旋转或平移之后.

效果如下图所示:

转载于:http://blog.csdn.net/mduanfire/article/details/51674098

 

 为什么要做变形,或者说数据增强。从这个网站可以看出 http://scs.ryerson.ca/~aharley/vis/conv/           手写字符稍微变形点,就有可能识别出错,因此数据增强可以生成一些变形的数据,让网络提前适应

 

 

 1 # -*- coding: utf-8 -*-
 2 __author__ = 'Administrator'
 3 
 4 # import packages
 5 from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
 6 
 7 datagen = ImageDataGenerator(
 8         rotation_range=0.2,
 9         width_shift_range=0.2,
10         height_shift_range=0.2,
11         shear_range=0.2,
12         zoom_range=0.2,
13         horizontal_flip=True,
14         fill_mode='nearest')
15 
16 for k in range(33):
17     numstr = "{0:d}".format(k);
18     filename='C:\Users\Administrator\Desktop\bad\'+numstr+'.jpg';
19     ufilename = unicode(filename , "utf8")
20     img = load_img(ufilename)  # this is a PIL image, please replace to your own file path
21     x = img_to_array(img)  # this is a Numpy array with shape (3, 150, 150)
22     x = x.reshape((1,) + x.shape)  # this is a Numpy array with shape (1, 3, 150, 150)
23 
24     # the .flow() command below generates batches of randomly transformed images
25     # and saves the results to the `preview/` directory
26 
27     i = 0
28 
29     for batch in datagen.flow(x,
30                               batch_size=1,
31                               save_to_dir='C:\Users\Administrator\Desktop\dataA\',#生成后的图像保存路径
32                               save_prefix=numstr,
33                               save_format='jpg'):
34         i += 1
35         if i > 20:
36             break  # otherwise the generator would loop indefinitely
37 end

View Code

 

 1 # -*- coding: utf-8 -*-
 2 __author__ = 'Administrator'
 3 
 4 # import packages
 5 from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
 6 
 7 datagen = ImageDataGenerator(
 8         rotation_range=10,
 9         width_shift_range=0.2,
10         height_shift_range=0.2,
11         rescale=1./255,
12         shear_range=0.2,
13         zoom_range=0.2,
14         horizontal_flip=True,
15         fill_mode='nearest')
16 import os
17 
18 import sys
19 reload(sys)
20 sys.setdefaultencoding('utf8')
21 
22 ufilename = unicode("C:\Users\Administrator\Desktop\测试" , "utf8")
23 
24 for filename in os.listdir(ufilename):              #listdir的参数是文件夹的路径
25     print ( filename)                                  #此时的filename是文件夹中文件的名称
26     pathname='C:\Users\Administrator\Desktop\测试\'+filename;
27     #ufilename = unicode(pathname , "utf8")
28     img = load_img(pathname)  # this is a PIL image, please replace to your own file path
29     x = img_to_array(img)  # this is a Numpy array with shape (3, 150, 150)
30     x = x.reshape((1,) + x.shape)  # this is a Numpy array with shape (1, 3, 150, 150)
31     # the .flow() command below generates batches of randomly transformed images
32     # and saves the results to the `preview/` directory
33     i = 0
34     for batch in datagen.flow(x,
35                               batch_size=1,
36                               save_to_dir='C:\Users\Administrator\Desktop\result\',#生成后的图像保存路径
37                               save_prefix=filename,
38                               save_format='jpg'):
39         i += 1
40         if i > 100:
41             break  # otherwise the generator would loop indefinitely
42 
43 
44 # datagen = ImageDataGenerator(
45 #         rotation_range=0.2,
46 #         width_shift_range=0.2,
47 #         height_shift_range=0.2,
48 #         rescale=1./255,
49 #         shear_range=0.1,
50 #         zoom_range=0.4,
51 #         horizontal_flip=True,
52 #         fill_mode='nearest')
53 #
54 # ufilename = unicode("C:\Users\Administrator\Desktop\训练" , "utf8")
55 # for filename in os.listdir(ufilename):              #listdir的参数是文件夹的路径
56 #     print ( filename)                                  #此时的filename是文件夹中文件的名称
57 #     pathname='C:\Users\Administrator\Desktop\训练\'+filename;
58 #    # ufilename = unicode(pathname , "utf8")
59 #     img = load_img(pathname)  # this is a PIL image, please replace to your own file path
60 #     x = img_to_array(img)  # this is a Numpy array with shape (3, 150, 150)
61 #     x = x.reshape((1,) + x.shape)  # this is a Numpy array with shape (1, 3, 150, 150)
62 #
63 #     # the .flow() command below generates batches of randomly transformed images
64 #     # and saves the results to the `preview/` directory
65 #
66 #     i = 0
67 #
68 #     for batch in datagen.flow(x,
69 #                               batch_size=1,
70 #                               save_to_dir='C:\Users\Administrator\Desktop\result\',#生成后的图像保存路径
71 #                               save_prefix=filename,
72 #                               save_format='jpg'):
73 #         i += 1
74 #         if i > 100:
75 #             break  # otherwise the generator would loop indefinitely

View Code

 

https://github.com/mdbloice/Augmentor

 

本文链接:http://task.lmcjl.com/news/12536.html

展开阅读全文