Master Real-Time Image Augmentation with Keras ImageDataGenerator

Data augmentation expands a dataset by applying transformations such as cropping, padding, and flipping, making models more robust to small variations and reducing overfitting. Storing all augmented images in memory is inefficient, which is why Keras’s ImageDataGenerator (available via tensorflow.keras) generates augmented batches in real time.

Quick Start

The generator creates batches of tensors with the same dimensions as the input images. The following helper script visualizes the effects of each augmentation parameter.

from tensorflow.keras.preprocessing.image import ImageDataGenerator
from matplotlib.pyplot import imread, imshow, subplots, show

def plot(data_generator):
    """Plot 4 images generated by an ImageDataGenerator instance."""
    data_generator.fit(images)
    image_iterator = data_generator.flow(images)
    fig, rows = subplots(nrows=1, ncols=4, figsize=(18,18))
    for row in rows:
        row.imshow(image_iterator.next()[0].astype('int'))
        row.axis('off')
    show()

image = imread("image.jpeg")
images = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
imshow(images[0])
show()

1. Rotation

data_generator = ImageDataGenerator(rotation_range=90)
plot(data_generator)

Images are randomly rotated within ±90 degrees.

2. Width Shifting

data_generator = ImageDataGenerator(width_shift_range=0.3)
plot(data_generator)

width_shift_range

(0.0–1.0) defines the maximum fraction of total width by which an image can shift left or right.

3. Height Shifting

data_generator = ImageDataGenerator(height_shift_range=0.3)
plot(data_generator)

Similar to width shifting, but moves the image vertically.

4. Brightness

data_generator = ImageDataGenerator(brightness_range=(0.1, 0.9))
plot(data_generator)

brightness_range

selects a random brightness offset; 0.0 means no brightness, 1.0 the maximum.

5. Shear Intensity

data_generator = ImageDataGenerator(shear_range=45.0)
plot(data_generator)

shear_range

(in degrees) tilts the image shape, creating a stretching effect distinct from rotation.

6. Zoom

data_generator = ImageDataGenerator(zoom_range=[0.5, 1.5])
plot(data_generator)

Values <1.0 zoom in, >1.0 zoom out, applying random scaling.

7. Channel Shift

data_generator = ImageDataGenerator(channel_shift_range=150.0)
plot(data_generator)

Randomly adds a value (chosen from the specified range) to each channel.

8. Horizontal Flip

data_generator = ImageDataGenerator(horizontal_flip=True)
plot(data_generator)

Generates horizontally mirrored images.

9. Vertical Flip

data_generator = ImageDataGenerator(vertical_flip=True)
plot(data_generator)

Generates vertically mirrored images.

10. Fill Modes

When transformations create empty pixels, four fill strategies are available:

nearest : Replicates the nearest pixel value.

reflect : Mirrors the border pixels.

wrap : Wraps around the image edges.

constant : Fills with a constant value defined by cval.

# Example using constant fill mode
data_generator = ImageDataGenerator(width_shift_range=0.3, fill_mode='constant', cval=190)
plot(data_generator)

Additional Options

Other useful parameters include featurewise/ samplewise centering and normalization ( featurewise_center, samplewise_center, featurewise_std_normalization, samplewise_std_normalization) and rescale to multiply all pixel values. A custom preprocessing_function can be supplied for user‑defined image processing.

These capabilities enable efficient, on‑the‑fly augmentation for training robust deep‑learning models.