Data augmentation using shear and rotate within a range.

shear_and_rotate.py

def shear_and_rotate(image_path, img_w, img_h):
    from skimage.io import imread, imsave
    from skimage.transform import resize
    from skimage import transform as transf
    import numpy as np
    from skimage.filters import threshold_otsu
    from scipy.ndimage.interpolation import rotate
    import math

    image = imread(image_path)

    augmented_images = []

    # Resize the image
    resized_ori = resize(image, (img_h, img_w))
    
    # Binarize the image using Otsu thresholding
    thresh = threshold_otsu(resized_ori)
    resized_ori = resized_ori > thresh
    resized_ori = resized_ori.astype(int)
    #resized_ori = np.transpose(resized_ori)
    augmented_images.append(resized_ori)

    for rotate_angle in np.arange(-5, 7, 2):
        rotated = rotate(image, rotate_angle, reshape=True, mode='constant', cval=1.)

        for shear_angle in np.arange(-0.5, 0.7, 0.2):
            affine_tf = transf.AffineTransform(shear=math.radians(shear_angle))
            sheared = transf.warp(rotated, affine_tf, mode='constant', cval=1.)

            resized = resize(sheared, (img_h, img_w))
            thresh = threshold_otsu(resized)
            resized = resized > thresh
            resized = resized.astype(int)

            #resized = np.transpose(resized) # For CNN-RNN only

            augmented_images.append(resized)

    augmented_images = np.array(augmented_images) # 36 additional images generated for each input image

    return augmented_images