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:
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
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Ixy=[IRxy,IGxy,IBxy]T进行加上如下的变化:
[p1,p2,p3][α1λ1,α2λ2,α3λ3]T
其中:αi是满足均值为0,方差为0.1的随机变量.
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:
# -*- coding: utf-8 -*-
__author__ = 'Administrator' # import packages
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img datagen = ImageDataGenerator(
rotation_range=0.2,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest') img = load_img('C:\Users\Administrator\Desktop\dataA\lena.jpg') # this is a PIL image, please replace to your own file path
x = img_to_array(img) # this is a Numpy array with shape (3, 150, 150)
x = x.reshape((1,) + x.shape) # this is a Numpy array with shape (1, 3, 150, 150) # the .flow() command below generates batches of randomly transformed images
# and saves the results to the `preview/` directory i = 0
for batch in datagen.flow(x,
batch_size=1,
save_to_dir='C:\Users\Administrator\Desktop\dataA\pre',#生成后的图像保存路径
save_prefix='lena',
save_format='jpg'):
i += 1
if i > 20: #这个20指出要扩增多少个数据
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
# -*- coding: utf-8 -*-
__author__ = 'Administrator' # import packages
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img datagen = ImageDataGenerator(
rotation_range=0.2,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest') for k in range(33):
numstr = "{0:d}".format(k);
filename='C:\\Users\\Administrator\\Desktop\\bad\\'+numstr+'.jpg';
ufilename = unicode(filename , "utf8")
img = load_img(ufilename) # this is a PIL image, please replace to your own file path
x = img_to_array(img) # this is a Numpy array with shape (3, 150, 150)
x = x.reshape((1,) + x.shape) # this is a Numpy array with shape (1, 3, 150, 150) # the .flow() command below generates batches of randomly transformed images
# and saves the results to the `preview/` directory i = 0 for batch in datagen.flow(x,
batch_size=1,
save_to_dir='C:\\Users\\Administrator\\Desktop\\dataA\\',#生成后的图像保存路径
save_prefix=numstr,
save_format='jpg'):
i += 1
if i > 20:
break # otherwise the generator would loop indefinitely
end
# -*- coding: utf- -*-
__author__ = 'Administrator' # import packages
from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img datagen = ImageDataGenerator(
rotation_range=,
width_shift_range=0.2,
height_shift_range=0.2,
rescale=./,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
import os import sys
reload(sys)
sys.setdefaultencoding('utf8') ufilename = unicode("C:\\Users\\Administrator\\Desktop\\测试" , "utf8") for filename in os.listdir(ufilename): #listdir的参数是文件夹的路径
print ( filename) #此时的filename是文件夹中文件的名称
pathname='C:\\Users\\Administrator\\Desktop\\测试\\'+filename;
#ufilename = unicode(pathname , "utf8")
img = load_img(pathname) # this is a PIL image, please replace to your own file path
x = img_to_array(img) # this is a Numpy array with shape (, , )
x = x.reshape((,) + x.shape) # this is a Numpy array with shape (, , , )
# the .flow() command below generates batches of randomly transformed images
# and saves the results to the `preview/` directory
i =
for batch in datagen.flow(x,
batch_size=,
save_to_dir='C:\\Users\\Administrator\\Desktop\\result\\',#生成后的图像保存路径
save_prefix=filename,
save_format='jpg'):
i +=
if i > :
break # otherwise the generator would loop indefinitely # datagen = ImageDataGenerator(
# rotation_range=0.2,
# width_shift_range=0.2,
# height_shift_range=0.2,
# rescale=./,
# shear_range=0.1,
# zoom_range=0.4,
# horizontal_flip=True,
# fill_mode='nearest')
#
# ufilename = unicode("C:\\Users\\Administrator\\Desktop\\训练" , "utf8")
# for filename in os.listdir(ufilename): #listdir的参数是文件夹的路径
# print ( filename) #此时的filename是文件夹中文件的名称
# pathname='C:\\Users\\Administrator\\Desktop\\训练\\'+filename;
# # ufilename = unicode(pathname , "utf8")
# img = load_img(pathname) # this is a PIL image, please replace to your own file path
# x = img_to_array(img) # this is a Numpy array with shape (, , )
# x = x.reshape((,) + x.shape) # this is a Numpy array with shape (, , , )
#
# # the .flow() command below generates batches of randomly transformed images
# # and saves the results to the `preview/` directory
#
# i =
#
# for batch in datagen.flow(x,
# batch_size=,
# save_to_dir='C:\\Users\\Administrator\\Desktop\\result\\',#生成后的图像保存路径
# save_prefix=filename,
# save_format='jpg'):
# i +=
# if i > :
# break # otherwise the generator would loop indefinitely
https://github.com/mdbloice/Augmentor