基于Tensorflow深度学习框架下的图像增强(一)

深度学习最依赖的就是数据量了，同样的一张照片轻松一变数据量就翻倍了

所有可能的变化，比如说平移或者按角度的旋转,放大缩小综合的随机执行这些变换

导入基础python包

import matplotlib.pyplot as plt
from PIL import Image
%matplotlib inline
from keras.preprocessing import image #函数在Image模块下，图像变化现成的框架工具包，平移，放缩，旋转
# keras预处理模块下有一个image对图像数据进行处理操作
import keras.backend as K
import os
import glob
import numpy as np

展示输入数据

def print_result(path):
    name_list = glob.glob(path)
    fig = plt.figure(figsize=(12,16)) #定义了图像的大小
    for i in range(3):
        img = Image.open(name_list[i])
        sub_img = fig.add_subplot(131+i) #加了一个子图
        sub_img.imshow(img)

img_path = '/Users/mac/Desktop/chest_xray/img/x/*'
in_path = '/Users/mac/Desktop/chest_xray/img/'
out_path = '/Users/mac/Desktop/chest_xray/output/'
name_list = glob.glob(img_path)
name_list

['/Users/mac/Desktop/chest_xray/img/x/person117_bacteria_553.jpeg',
 '/Users/mac/Desktop/chest_xray/img/x/person117_bacteria_556.jpeg',
 '/Users/mac/Desktop/chest_xray/img/x/person117_bacteria_557.jpeg']

print_result(img_path)

指定target_size后的图像都变为相同大小好送入神经网络

datagen = image.ImageDataGenerator()
gen_data = datagen.flow_from_directory(in_path,batch_size=1,shuffle=False,
                                      save_to_dir = out_path+'resize',
                                       save_prefix = 'gen',target_size = (224,224))
#save_prefix = 'gen'加前缀方便自己识别
#把三张图像都执行一次操作
for i in range(3):
    gen_data.next()
print_result(out_path+'resize/*')

 

角度变换

datagen = image.ImageDataGenerator(rotation_range=45)
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'rotation_range',
                                       save_prefix='gen',target_size=(224,224))for i in range(3):
    gen_data.next()
print_result(out_path+'rotation_range/*')

 

平移操作

datagen = image.ImageDataGenerator(width_shift_range=0.3,height_shift_range=0.3)
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'shift',
                                       save_prefix='gen',target_size=(224,224))
for i in range(3):
    gen_data.next()
print_result(out_path+'shift/*')

 

datagen = image.ImageDataGenerator(width_shift_range=-0.3,height_shift_range=0.3)
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'shift2',
                                       save_prefix='gen',target_size=(224,224))for i in range(3):
    gen_data.next()
print_result(out_path+'shift2/*')

 

缩放

datagen = image.ImageDataGenerator(zoom_range=0.5)
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'zoom',
                                       save_prefix='gen',target_size=(224,224))
for i in range(3):
    gen_data.next()
print_result(out_path+'zoom/*')

 

 channel_shift 对颜色通道值的一个变换

datagen = image.ImageDataGenerator(channel_shift_range=15)
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'channel',
                                       save_prefix='gen',target_size=(224,224))
for i in range(3):
    gen_data.next()
print_result(out_path+'channel/*')

翻转

datagen = image.ImageDataGenerator(horizontal_flip=True)
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'horizontal',
                                       save_prefix='gen',target_size=(224,224))
for i in range(3):
    gen_data.next()
print_result(out_path+'horizontal/*')

 

可能我们看图像的话是没有什么太大差别的，但是我们把这些图像转换为向量就可以看见不同变化的差别。

gen_data = datagen.flow_from_directory(in_path,batch_size=1,
                                      shuffle=False,save_to_dir=out_path+'horizontal',
                                      save_prefix='gen',target_size=(224,224))
gen_data.next()

 

Found 3 images belonging to 1 classes.
(array([[[[ 44.,  44.,  44.],
          [  9.,   9.,   9.],
          [ 87.,  87.,  87.],
          ...,
          [ 13.,  13.,  13.],
          [ 12.,  12.,  12.],
          [ 14.,  14.,  14.]],
 
         [[ 39.,  39.,  39.],
          [ 47.,  47.,  47.],
          [202., 202., 202.],
          ...,
          [ 13.,  13.,  13.],
          [ 12.,  12.,  12.],
          [ 14.,  14.,  14.]],
 
         [[ 20.,  20.,  20.],
          [184., 184., 184.],
          [193., 193., 193.],
          ...,
          [ 13.,  13.,  13.],
          [ 12.,  12.,  12.],
          [ 14.,  14.,  14.]],
 
         ...,
 
         [[ 15.,  15.,  15.],
          [ 15.,  15.,  15.],
          [ 15.,  15.,  15.],
          ...,
          [ 19.,  19.,  19.],
          [ 19.,  19.,  19.],
          [ 19.,  19.,  19.]],
 
         [[ 15.,  15.,  15.],
          [ 15.,  15.,  15.],
          [ 15.,  15.,  15.],
          ...,
          [ 19.,  19.,  19.],
          [ 19.,  19.,  19.],
          [ 19.,  19.,  19.]],
 
         [[ 15.,  15.,  15.],
          [ 15.,  15.,  15.],
          [ 15.,  15.,  15.],
          ...,
          [ 19.,  19.,  19.],
          [ 19.,  19.,  19.],
          [ 19.,  19.,  19.]]]], dtype=float32),
 array([[1.]], dtype=float32))

rescale(归一化)

datagen = image.ImageDataGenerator(rescale = 1/255) #都除以255
gen = image.ImageDataGenerator()
data = gen.flow_from_directory(in_path,batch_size=1,class_mode=None,shuffle = True, target_size = (224,224))
np_data = np.concatenate([data.next() for i in range(data.n)])
datagen.fit(np_data)
gen_data = datagen.flow_from_directory(in_path,batch_size = 1, shuffle = False, 
                                       save_to_dir = out_path+'rescale',
                                       save_prefix='gen',target_size=(224,224))
for i in range(3):
    gen_data.next()
print_result(out_path+'rescale/*')

gen_data = datagen.flow_from_directory(in_path,batch_size=1, #每篇论文必须要去做的，把像素压缩到0-1之间
                                      shuffle=False,save_to_dir=out_path+'rescale',
                                      save_prefix='gen',target_size=(224,224))
gen_data.next()

 

(array([[[[0.05490196, 0.05490196, 0.05490196],
          [0.04705883, 0.04705883, 0.04705883],
          [0.0509804 , 0.0509804 , 0.0509804 ],
          ...,
          [0.34117648, 0.34117648, 0.34117648],
          [0.03529412, 0.03529412, 0.03529412],
          [0.17254902, 0.17254902, 0.17254902]],
 
         [[0.05490196, 0.05490196, 0.05490196],
          [0.04705883, 0.04705883, 0.04705883],
          [0.0509804 , 0.0509804 , 0.0509804 ],
          ...,
          [0.79215693, 0.79215693, 0.79215693],
          [0.18431373, 0.18431373, 0.18431373],
          [0.15294118, 0.15294118, 0.15294118]],
 
         [[0.05490196, 0.05490196, 0.05490196],
          [0.04705883, 0.04705883, 0.04705883],
          [0.0509804 , 0.0509804 , 0.0509804 ],
          ...,
          [0.7568628 , 0.7568628 , 0.7568628 ],
          [0.72156864, 0.72156864, 0.72156864],
          [0.07843138, 0.07843138, 0.07843138]],
 
         ...,
 
         [[0.07450981, 0.07450981, 0.07450981],
          [0.07450981, 0.07450981, 0.07450981],
          [0.07450981, 0.07450981, 0.07450981],
          ...,
          [0.05882353, 0.05882353, 0.05882353],
          [0.05882353, 0.05882353, 0.05882353],
          [0.05882353, 0.05882353, 0.05882353]],
 
         [[0.07450981, 0.07450981, 0.07450981],
          [0.07450981, 0.07450981, 0.07450981],
          [0.07450981, 0.07450981, 0.07450981],
          ...,
          [0.05882353, 0.05882353, 0.05882353],
          [0.05882353, 0.05882353, 0.05882353],
          [0.05882353, 0.05882353, 0.05882353]],
 
         [[0.07450981, 0.07450981, 0.07450981],
          [0.07450981, 0.07450981, 0.07450981],
          [0.07450981, 0.07450981, 0.07450981],
          ...,
          [0.05882353, 0.05882353, 0.05882353],
          [0.05882353, 0.05882353, 0.05882353],
          [0.05882353, 0.05882353, 0.05882353]]]], dtype=float32),
 array([[1.]], dtype=float32))

填充方法:

'constant':kkkkkkkk|abcd|kkkkkkkk(cval=k)

'nearest':aaaaaaaa|abcd|dddddddd

'reflect':abcddcba|abcd|dcbaabcd

'wrap':abcdabcd|abcd|abcdabcd