Tensorflow: Crop largest central square region of image

Question

My network takes images of size 100 x 100 pixels. Therefore I have to resize the images of my dataset which are of different size. I want to be able to extract the largest central square region from a given image and then resize it to 100 x 100.

To be more precisely, let's say an image has a width of 200 pixels and a height of 50 pixels. Then I want to extract the largest central square region which is in this example 50 x 50 followed by resizing the image to 100 x 100 pixels.

What is the right way to do that using Tensorflow? Right now I am using tf.image.resize_images() which distorts the image and I want to get rid of that.

Answer 1

Sounds like crop_to_bounding_box is doing what you need:

import tensorflow as tf

def crop_center(image):
    h, w = image.shape[-3], image.shape[-2]
    if h > w:
        cropped_image = tf.image.crop_to_bounding_box(image, (h - w) // 2, 0, w, w)
    else:
        cropped_image = tf.image.crop_to_bounding_box(image, 0, (w - h) // 2, h, h)
    return tf.image.resize_images(cropped_image, (100, 100))

Answer 2

I think this does what you want:

import tensorflow as tf

def crop_center_and_resize(img, size):
    s = tf.shape(img)
    w, h = s[0], s[1]
    c = tf.minimum(w, h)
    w_start = (w - c) // 2
    h_start = (h - c) // 2
    center = img[w_start:w_start + c, h_start:h_start + c]
    return tf.image.resize_images(img, [size, size])

print(crop_center_and_resize(tf.zeros((80, 50, 3)), 100))
# Tensor("resize_images/Squeeze:0", shape=(100, 100, 3), dtype=float32)

There is also tf.image.crop_and_resize, which can do both things in one go, but you have to use normalized image coordinates with that:

import tensorflow as tf

def crop_center_and_resize(img, size):
    s = tf.shape(img)
    w, h = s[0], s[1]
    c = tf.minimum(w, h)
    wn, hn = h / c, w / c
    result = tf.image.crop_and_resize(tf.expand_dims(img, 0),
                                      [[(1 - wn) / 2, (1 - hn) / 2, wn, hn]],
                                      [0], [size, size])
    return tf.squeeze(result, 0)