Data pre-Processing and data augmentation

What is data augmentation ?

Term “augmentation”- The action of making or becoming greater in size or amount.

Data augmentation mean increasing the amount of Data by applying random transformation, so that our model would never see exact same picture twice.

Why Data augmentation is needed ?

• To increase the performance of deep learning neural networks often improves with the amount of data available.
• To prevent overfitting and helps the model generalize better

Data augmentation techniques

1. Flip: Flipping images on horizontal or vetical axis

2. Rotation: Rotate an image with certain degree

3. Crop: Randomly, crop a section from a given image and resize

4. Add Noise: Adding Gaussian noise to a given image

5. Color Jittering: Random color manipulation

Data augmentation using Keras

In keras this can be done via the keras.preprocessing.image.ImageDataGenerator class. This class allows

• configure random transformations and normalization operations to be done on your image data during training.

Let us look at an example:

from keras.preprocessing.image import ImageDataGenerator, load_img, img_to_array
import os
datagen= ImageDataGenerator(
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
rescale=1./255,
shear_range=0.2,
horizontal_flip=True,
fill_mode='nearest'
)

#creating a directory to save results

if not os.path.exists('data_aug'):
    os.makedirs('data_aug')

img=load_img('C:/Users/Asus/Desktop/car.jpg')
print(img)
x=img_to_array(img) #this is a numpy img_to_array
x=x.reshape((1,) + x.shape)

# the .flow() command below generates batches of randomly transformed images
# saves the results to the 'data_aug' directory

i=1
for batch in datagen.flow(x, batch_size=1,
                          save_to_dir='data_aug', save_prefix='car', save_format='jpeg'):


    i +=1
    if i > 20:

        break #otherwise the generator would loop indefinitely

Here’s what we get – this is our data strategy looks like: