Data augmentation using ImageDataGenerator

Question

I'm trying to do a simple data augmentation on my data using ImageDataGenerator. However, my generator is giving my distorted images due to the zoom parameter. I would like the random zoom to only zoom in on my data, and when it zooms in to apply the same zoom on the width and height (to avoid distorted image outputs).

gen = ImageDataGenerator(featurewise_center=False, featurewise_std_normalization=False,
                         samplewise_center = True, samplewise_std_normalization=True,
                         rotation_range=180, width_shift_range=0, height_shift_range=0,
                         shear_range=0, zoom_range=0.4, fill_mode='reflect',
                        horizontal_flip=True, rescale=1./255) 

For now, the resulting images are below.

original image output

Answer 1

What you're looking for is random cropping:

def random_crop(image):
    height, width = image.shape[:2]
    random_array = numpy.random.random(size=4);
    w = int((width*0.5) * (1+random_array[0]*0.5))
    h = int((height*0.5) * (1+random_array[1]*0.5))
    x = int(random_array[2] * (width-w))
    y = int(random_array[3] * (height-h))

    image_crop = image[y:h+y, x:w+x, 0:3]
    image_crop = resize(image_crop, image.shape)
    return image_crop

# Data generator 
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',
                             preprocessing_function=random_crop)

Source here .