Data enhancement

Catalog

1. Implementation method of data enhancement

2. Data enhancement in tf In the implementation of

2.1  adopt ImageDataGenerator Enhance the data

---

1. Implementation method of data enhancement

（1） normalization
    Normalize the data to a certain range , Make the network easier to learn .

（2） Image transformation
  Flip , The tensile , tailoring , deformation （ Rich data ）

（3） Color change
  Contrast , brightness

（4） Multiscale clipping

2. Data enhancement in tf In the implementation of

adopt keras.preprocessing.image.ImageDataGenerator Interface to implement data enhancement

rescale = 1./255, # Zoom all pixels to 0——1, Because the values of all pixels are 0——255 Between .

rotation_range = 40, # A way of image enhancement , such as ratation_range = 40, Indicates that the random rotation angle of the image is -40——40 Between degrees .

width_shift_range = 0.2, # Displacement in the horizontal direction , Less than 1, It is the ratio of displacement ; Greater than 1, Is the value of the moving pixel .width_shift_range = 0.2, Indicates random translation of the image 0——20%

height_shift_range = 0.2, # Displacement in the vertical direction

shear_range = 0.2, # Shear strength

zoom_range = 0.2, # Scaling strength

horizontal_flip = True, # Whether to do random inversion in the horizontal direction

fill_mode = 'nearest') # When processing pictures , Use when pixels need to be filled . such as , Enlarge the image , Some pixels need to be filled .

2.1  adopt ImageDataGenerator Enhance the data

#  Training data 
train_datagen= keras.preprocessing.image.ImageDataGenerator(
    rescale = 1./255, #  Zoom all pixels to 0——1, Because the values of all pixels are 0——255 Between .
    rotation_range = 40, #  A way of image enhancement , such as  ratation_range = 40,  Indicates that the random rotation angle of the image is -40——40 Between degrees .
    width_shift_range = 0.2, #  Displacement in the horizontal direction , Less than 1, It is the ratio of displacement ; Greater than 1, Is the value of the moving pixel .width_shift_range = 0.2, Indicates random translation of the image 0——20%
    height_shift_range = 0.2, #  Displacement in the vertical direction 
    shear_range = 0.2, #  Shear strength 
    zoom_range = 0.2, #  Scaling strength 
    horizontal_flip = True, #  Whether to do random inversion in the horizontal direction 
    fill_mode = 'nearest') #  When processing pictures , Use when pixels need to be filled . such as , Enlarge the image , Some pixels need to be filled .


train_generator = train_datagen.flow_from_directory(
    train_dir, 
    target_size = (height, width),
    batch_size = batch_size,
    seed = 7, 
    shuffle = True,
    class_mode = "categorical")

#  Validation data 
valid_datagen= keras.preprocessing.image.ImageDataGenerator(
    rescale = 1./255 #  Zoom all pixels to 0——1, Because the values of all pixels are 0——255 Between .
    )


valid_generator = train_datagen.flow_from_directory(
    valid_dir, 
    target_size = (height, width),
    batch_size = batch_size,
    seed = 7, 
    shuffle = False,
    class_mode = "categorical")

For more codes, see ： This code is based on convolutional neural network .