useTensorFlow rotates images randomly

When using deep learning to train images, random rotation of the images helps to improve the model generalization ability. However, before doing preprocessing such as rotation, the image is first rotated and saved locally, and then the model is input for training. This process will increase the workload. If there are many pictures, the generated rotated image will take up More space. Randomly rotate the image directly during the training process, which can effectively improve work efficiency and save hard disk space.

Use TensorFlow to randomly rotate the image as follows:

TensorFlow version is 1.13.1

#-*- coding:utf-8 -*-
'''
    Example of random rotation of images using TensorFlow
'''

import tensorflow as tf
import numpy as np
import cv2
import matplotlib.pyplot as plt


img = cv2.imread('tf.jpg')
img = cv2.resize(img,(220,220))
img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)

def tf_rotate(input_image, min_angle = -np.pi/2, max_angle = np.pi/2):
    '''
         TensorFlow rotates images randomly
         :param input_image: image input
         :param min_angle: minimum rotation angle
         :param max_angle: maximum rotation angle
         :return: rotated image
    '''
    distorted_image = tf.expand_dims(input_image, 0)
    random_angles = tf.random.uniform(shape=(tf.shape(distorted_image)[0],), minval = min_angle , maxval = max_angle)
    distorted_image = tf.contrib.image.transform(
        distorted_image,
        tf.contrib.image.angles_to_projective_transforms(
            random_angles, tf.cast(tf.shape(distorted_image)[1], tf.float32), tf.cast(tf.shape(distorted_image)[2], tf.float32)
        ))
    rotate_image = tf.squeeze(distorted_image, [0])
    return rotate_image

global_init = tf.global_variables_initializer()
with tf.Session() as sess:
    init = tf.initialize_local_variables()
    sess.run([init, global_init])
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)
    image = tf.placeholder(shape=(220, 220, 3), dtype=tf.float32)

    rotate_image = tf_rotate(image, -np.pi/2, np.pi/2)
    output = sess.run(rotate_image, feed_dict={image:img})
    # print('output:',output)
    plt.imshow(output.astype('uint8'))
    plt.title('rotate image')
    plt.show()

The results are as follows:

Original image:

Randomly rotated picture: